Merge remote-tracking branch 'origin/master' into set-default-subgroup-for-intel

.github/labeler.yml

@@ -75,21 +75,13 @@ android:
             - examples/llama.android/**
 server/webui:
     - changed-files:
-        - all:
-            - any-glob-to-any-file:
-                - tools/server/webui/**
-                - tools/server/public/**
-            - all-globs-to-all-files:
-                - '!tools/server/webui/**'
-                - '!tools/server/public/**'
+        - any-glob-to-any-file:
+            - tools/server/webui/**
+            - tools/server/public/**
 server:
     - changed-files:
-        - all:
-            - any-glob-to-any-file:
-                - tools/server/**
-            - all-globs-to-all-files:
-                - '!tools/server/webui/**'
-                - '!tools/server/public/**'
+        - any-glob-to-any-file:
+            - tools/server/**
 
 
 

common/arg.cpp

@@ -2348,19 +2348,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_N_GPU_LAYERS"));
     add_opt(common_arg(
-        {"-sm", "--split-mode"}, "{none,layer,row}",
+        {"-sm", "--split-mode"}, "{none,layer,row,tensor}",
         "how to split the model across multiple GPUs, one of:\n"
         "- none: use one GPU only\n"
-        "- layer (default): split layers and KV across GPUs\n"
-        "- row: split rows across GPUs",
+        "- layer (default): split layers and KV across GPUs (pipelined)\n"
+        "- row: split weight across GPUs by rows (parallelized)\n"
+        "- tensor: split weights and KV across GPUs (parallelized)",
         [](common_params & params, const std::string & value) {
-            std::string arg_next = value;
-            if (arg_next == "none") {
+            if (value == "none") {
                 params.split_mode = LLAMA_SPLIT_MODE_NONE;
-            } else if (arg_next == "layer") {
+            } else if (value == "layer") {
                 params.split_mode = LLAMA_SPLIT_MODE_LAYER;
-            } else if (arg_next == "row") {
+            } else if (value == "row") {
                 params.split_mode = LLAMA_SPLIT_MODE_ROW;
+            } else if (value == "tensor") {
+                params.split_mode = LLAMA_SPLIT_MODE_TENSOR;
             } else {
                 throw std::invalid_argument("invalid value");
             }

common/chat-auto-parser-generator.cpp

@@ -332,58 +332,36 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
     const auto & inputs      = ctx.inputs;
     bool         force_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
 
+    auto until_suffix = p.rule("until-suffix", p.until(arguments.value_suffix));
+
     common_peg_parser tool_choice = p.choice();
 
     foreach_function(inputs.tools, [&](const json & tool) {
         const auto &          func       = tool.at("function");
         std::string           name       = func.at("name");
-        const auto &          params     = func.contains("parameters") ? func.at("parameters") : json::object();
+        auto                  params     = func.contains("parameters") ? func.at("parameters") : json::object();
         const auto &          properties = params.contains("properties") ? params.at("properties") : json::object();
+
         std::set<std::string> required;
+        if (params.contains("required")) {
+            params.at("required").get_to(required);
+        }
+
+        auto schema_info = common_schema_info();
+        schema_info.resolve_refs(params);
 
         // Build parser for each argument, separating required and optional
         std::vector<common_peg_parser> required_parsers;
         std::vector<common_peg_parser> optional_parsers;
         for (const auto & [param_name, param_schema] : properties.items()) {
-            bool        is_required = required.find(param_name) != required.end();
-            std::string type        = "object";
-            if (param_schema.contains("type")) {
-                const auto & type_obj = param_schema.at("type");
-                if (type_obj.is_string()) {
-                    type_obj.get_to(type);
-                } else if (type_obj.is_array()) {
-                    // Handle nullable types like ["string", "null"]
-                    for (const auto & t : type_obj) {
-                        if (t.is_string() && t.get<std::string>() != "null") {
-                            type = t.get<std::string>();
-                            break;
-                        }
-                    }
-                } else if (type_obj.is_object()) {
-                    if (type_obj.contains("type") && type_obj.at("type").is_string()) {
-                        type_obj.at("type").get_to(type);
-                    }
-                }
-            }
-            // Infer string type from enum values when type is unspecified
-            if (type == "object" && param_schema.contains("enum")) {
-                const auto & enum_vals = param_schema.at("enum");
-                if (enum_vals.is_array()) {
-                    for (const auto & v : enum_vals) {
-                        if (v.is_string()) {
-                            type = "string";
-                            break;
-                        }
-                    }
-                }
-            }
+            bool is_required = required.find(param_name) != required.end();
 
             auto arg =
                 p.tool_arg(p.tool_arg_open(arguments.name_prefix + p.tool_arg_name(p.literal(param_name)) +
                                            arguments.name_suffix) +
                            arguments.value_prefix +
-                           (type == "string" ?
-                                p.tool_arg_string_value(p.schema(p.until(arguments.value_suffix),
+                           (schema_info.resolves_to_string(param_schema) ?
+                                p.tool_arg_string_value(p.schema(until_suffix,
                                                                  "tool-" + name + "-arg-" + param_name + "-schema",
                                                                  param_schema, true)) :
                                 p.tool_arg_json_value(p.schema(
@@ -414,7 +392,7 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
             for (const auto & opt : optional_parsers) {
                 any_opt |= opt;
             }
-            args_seq = args_seq + p.repeat(p.space() + any_opt, 0, (int) optional_parsers.size());
+            args_seq = args_seq + p.repeat(p.space() + any_opt, 0, -1);
         }
 
         if (!arguments.start.empty()) {

common/chat.cpp

@@ -1124,7 +1124,7 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
             p.rule("gemma4-bool", p.json_bool());
             p.rule("gemma4-null", p.json_null());
             p.rule("gemma4-number", p.json_number());
-            p.rule("gemma4-dict-key", p.rule("gemma4-dict-key-name", p.until(":")) + p.literal(":"));
+            p.rule("gemma4-dict-key", p.rule("gemma4-dict-key-name", p.chars("[^:}]", 1, -1)) + p.literal(":"));
             p.rule("gemma4-dict-kv", p.ref("gemma4-dict-key") + p.space() + p.ref("gemma4-value"));
             p.rule("gemma4-dict", [&]() {
                 auto ws = p.space();

common/jinja/runtime.cpp

@@ -251,6 +251,23 @@ value binary_expression::execute_impl(context & ctx) {
         return res;
     }
 
+    // Python-style string repetition
+    // TODO: support array/tuple repetition (e.g., [1, 2] * 3 → [1, 2, 1, 2, 1, 2])
+    if (op.value == "*" &&
+            ((is_val<value_string>(left_val) && is_val<value_int>(right_val)) ||
+             (is_val<value_int>(left_val) && is_val<value_string>(right_val)))) {
+        const auto & str = is_val<value_string>(left_val) ? left_val->as_string() : right_val->as_string();
+        const int64_t repeat = is_val<value_int>(right_val) ? right_val->as_int() : left_val->as_int();
+        auto res = mk_val<value_string>();
+        if (repeat <= 0) {
+            return res;
+        }
+        for (int64_t i = 0; i < repeat; ++i) {
+            res->val_str = res->val_str.append(str);
+        }
+        return res;
+    }
+
     // String membership
     if (is_val<value_string>(left_val) && is_val<value_string>(right_val)) {
         // case: "a" in "abc"

common/jinja/value.cpp

@@ -1,4 +1,5 @@
 #include "runtime.h"
+#include "unicode.h"
 #include "value.h"
 
 // for converting from JSON to jinja values
@@ -154,6 +155,83 @@ static value test_compare_fn(const func_args & args) {
     return mk_val<value_bool>(value_compare(args.get_pos(0), args.get_pos(1), op));
 }
 
+static void append_codepoint_as_ascii_json_escape(std::string & out, uint32_t codepoint) {
+    auto append_u16 = [&out](uint32_t value) {
+        char buf[8];
+        snprintf(buf, sizeof(buf), "\\u%04x", static_cast<unsigned int>(value));
+        out += buf;
+    };
+
+    if (codepoint <= 0xFFFF) {
+        append_u16(codepoint);
+        return;
+    }
+
+    codepoint -= 0x10000;
+    append_u16(0xD800 + ((codepoint >> 10) & 0x3FF));
+    append_u16(0xDC00 + (codepoint & 0x3FF));
+}
+
+static std::string json_ensure_ascii_preserving_format(const std::string & json_str) {
+    std::string output;
+    output.reserve(json_str.size());
+
+    bool in_string = false;
+    bool escaped = false;
+
+    for (size_t pos = 0; pos < json_str.size();) {
+        const char ch = json_str[pos];
+        if (!in_string) {
+            output.push_back(ch);
+            if (ch == '"') {
+                in_string = true;
+            }
+            ++pos;
+            continue;
+        }
+
+        if (escaped) {
+            output.push_back(ch);
+            escaped = false;
+            ++pos;
+            continue;
+        }
+
+        if (ch == '\\') {
+            output.push_back(ch);
+            escaped = true;
+            ++pos;
+            continue;
+        }
+
+        if (ch == '"') {
+            output.push_back(ch);
+            in_string = false;
+            ++pos;
+            continue;
+        }
+
+        const unsigned char uch = static_cast<unsigned char>(ch);
+        if (uch < 0x80) {
+            output.push_back(ch);
+            ++pos;
+            continue;
+        }
+
+        auto parsed = common_parse_utf8_codepoint(json_str, pos);
+        if (parsed.status != utf8_parse_result::SUCCESS) {
+            output += "\\ufffd";
+            ++pos;
+            continue;
+        }
+
+        append_codepoint_as_ascii_json_escape(output, parsed.codepoint);
+        pos += parsed.bytes_consumed;
+    }
+
+    return output;
+}
+
 static value tojson(const func_args & args) {
     args.ensure_count(1, 5);
     value val_ascii      = args.get_kwarg_or_pos("ensure_ascii", 1);
@@ -169,16 +247,17 @@ static value tojson(const func_args & args) {
     if (is_val<value_int>(val_indent)) {
         indent = static_cast<int>(val_indent->as_int());
     }
-    if (val_ascii->as_bool()) { // undefined == false
-        throw not_implemented_exception("tojson ensure_ascii=true not implemented");
-    }
     if (val_sort->as_bool()) { // undefined == false
         throw not_implemented_exception("tojson sort_keys=true not implemented");
     }
+    const bool ensure_ascii = val_ascii->as_bool(); // undefined == false
     auto separators = (is_val<value_array>(val_separators) ? val_separators : mk_val<value_array>())->as_array();
     std::string item_sep = separators.size() > 0 ? separators[0]->as_string().str() : (indent < 0 ? ", " : ",");
     std::string key_sep = separators.size() > 1 ? separators[1]->as_string().str() : ": ";
     std::string json_str = value_to_json(args.get_pos(0), indent, item_sep, key_sep);
+    if (ensure_ascii) {
+        json_str = json_ensure_ascii_preserving_format(json_str);
+    }
     return mk_val<value_string>(json_str);
 }
 
@@ -460,6 +539,10 @@ const func_builtins & value_int_t::get_builtins() const {
             int64_t val = args.get_pos(0)->as_int();
             return mk_val<value_int>(val < 0 ? -val : val);
         }},
+        {"int", [](const func_args & args) -> value {
+            args.ensure_vals<value_int>();
+            return mk_val<value_int>(args.get_pos(0)->as_int());
+        }},
         {"float", [](const func_args & args) -> value {
             args.ensure_vals<value_int>();
             double val = static_cast<double>(args.get_pos(0)->as_int());
@@ -486,6 +569,10 @@ const func_builtins & value_float_t::get_builtins() const {
             int64_t val = static_cast<int64_t>(args.get_pos(0)->as_float());
             return mk_val<value_int>(val);
         }},
+        {"float", [](const func_args & args) -> value {
+            args.ensure_vals<value_float>();
+            return mk_val<value_float>(args.get_pos(0)->as_float());
+        }},
         {"safe", tojson},
         {"string", tojson},
         {"tojson", tojson},

convert_hf_to_gguf.py

@@ -1229,15 +1229,15 @@ class TextModel(ModelBase):
 
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
-        vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))
-        assert max(tokenizer.vocab.values()) < vocab_size
+        vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))  # ty: ignore[unresolved-attribute]
+        assert max(tokenizer.vocab.values()) < vocab_size  # ty: ignore[unresolved-attribute]
 
         tokpre = self.get_vocab_base_pre(tokenizer)
 
-        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
-        added_vocab = tokenizer.get_added_vocab()
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
+        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
 
-        added_tokens_decoder = tokenizer.added_tokens_decoder
+        added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
@@ -1250,7 +1250,7 @@ class TextModel(ModelBase):
                     # To avoid unexpected issues - we make sure to normalize non-normalized tokens
                     if not added_tokens_decoder[i].normalized:
                         previous_token = token
-                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
+                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[unresolved-attribute, invalid-assignment]
                         if previous_token != token:
                             logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
 
@@ -1583,13 +1583,13 @@ class TextModel(ModelBase):
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
         vocab_size = hparams["vocab_size"]
-        assert max(tokenizer.get_vocab().values()) < vocab_size
+        assert max(tokenizer.get_vocab().values()) < vocab_size  # ty: ignore[unresolved-attribute]
 
         tokpre = self.get_vocab_base_pre(tokenizer)
 
         merges = []
         vocab = {}
-        mergeable_ranks = tokenizer.mergeable_ranks
+        mergeable_ranks = tokenizer.mergeable_ranks  # ty: ignore[unresolved-attribute]
         for token, rank in mergeable_ranks.items():
             vocab[QwenModel.token_bytes_to_string(token)] = rank
             if len(token) == 1:
@@ -1599,7 +1599,7 @@ class TextModel(ModelBase):
             merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
 
         # for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined
-        added_vocab = tokenizer.special_tokens
+        added_vocab = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]
         reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **added_vocab}.items()}
 
         for i in range(vocab_size):
@@ -1622,10 +1622,10 @@ class TextModel(ModelBase):
         special_vocab.merges = merges
         # only add special tokens when they were not already loaded from config.json
         if len(special_vocab.special_token_ids) == 0:
-            special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"])
-            special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"])
+            special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+            special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
         # this one is usually not in config.json anyway
-        special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"])
+        special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def _set_vocab_sentencepiece(self, add_to_gguf=True):
@@ -1877,10 +1877,10 @@ class TextModel(ModelBase):
         self.gguf_writer.add_tokenizer_pre(tokpre)
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
-        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])
-        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def _set_vocab_glm(self):
@@ -1894,10 +1894,10 @@ class TextModel(ModelBase):
         self.gguf_writer.add_token_types(toktypes)
         # Special tokens
         # Note: Using <|endoftext|> (151329) for eot causes endless generation
-        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # 151331
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # 151336
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) # 151329
-        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # ty: ignore[unresolved-attribute]  # 151331
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # ty: ignore[unresolved-attribute]  # 151336
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]  # 151329
+        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # ty: ignore[unresolved-attribute]  # 151338
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def _set_vocab_interns1(self):
@@ -1906,16 +1906,16 @@ class TextModel(ModelBase):
 
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
-        vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())
+        vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())  # ty: ignore[unresolved-attribute]
         vocab_size = self.hparams.get("vocab_size", len(vocab))
         assert max(vocab.values()) < vocab_size
 
         tokpre = self.get_vocab_base_pre(tokenizer)
 
         reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab.items()}
-        added_vocab = tokenizer.get_added_vocab()
+        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
 
-        added_tokens_decoder = tokenizer.added_tokens_decoder
+        added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
@@ -1928,7 +1928,7 @@ class TextModel(ModelBase):
                     # To avoid unexpected issues - we make sure to normalize non-normalized tokens
                     if not added_tokens_decoder[i].normalized:
                         previous_token = token
-                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
+                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[unresolved-attribute, invalid-assignment]
                         if previous_token != token:
                             logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
 
@@ -2516,15 +2516,15 @@ class XverseModel(TextModel):
 
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(dir_model)
-        vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+        vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))  # ty: ignore[unresolved-attribute]
         # Since we are checking the maximum index, we need to ensure it's strictly less than vocab_size,
         # because vocab_size is the count of items, and indexes start at 0.
-        max_vocab_index = max(tokenizer.get_vocab().values())
+        max_vocab_index = max(tokenizer.get_vocab().values())  # ty: ignore[unresolved-attribute]
         if max_vocab_index >= vocab_size:
             raise ValueError("Vocabulary size exceeds expected maximum size.")
 
-        reverse_vocab: dict[int, str] = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
-        added_vocab = tokenizer.get_added_vocab()
+        reverse_vocab: dict[int, str] = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
+        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
 
         for token_id in range(vocab_size):
             token_text = reverse_vocab[token_id].encode('utf-8')
@@ -2535,7 +2535,7 @@ class XverseModel(TextModel):
             elif re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):
                 toktype = gguf.TokenType.BYTE  # special
             elif reverse_vocab[token_id] in added_vocab:
-                if tokenizer.added_tokens_decoder[token_id].special:
+                if tokenizer.added_tokens_decoder[token_id].special:  # ty: ignore[unresolved-attribute]
                     toktype = gguf.TokenType.CONTROL
                 else:
                     toktype = gguf.TokenType.USER_DEFINED
@@ -3752,7 +3752,7 @@ class QwenModel(TextModel):
 
     @staticmethod
     def token_bytes_to_string(b):
-        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
+        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode  # ty: ignore[unresolved-import]
         byte_encoder = bytes_to_unicode()
         return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])
 
@@ -3777,7 +3777,14 @@ class QwenModel(TextModel):
         self._set_vocab_qwen()
 
 
-@ModelBase.register("Qwen2Model", "Qwen2ForCausalLM", "Qwen2AudioForConditionalGeneration", "KORMoForCausalLM", "AudioFlamingo3ForConditionalGeneration")
+@ModelBase.register(
+    "Qwen2Model",
+    "Qwen2ForCausalLM",
+    "Qwen2AudioForConditionalGeneration",
+    "KORMoForCausalLM",
+    "AudioFlamingo3ForConditionalGeneration",
+    "DotsOCRForCausalLM",
+)
 class Qwen2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.QWEN2
 
@@ -3798,7 +3805,8 @@ class Qwen2Model(TextModel):
             name = name.replace("language_model.", "") # for InternVL
         if name.startswith("mlp") or name.startswith("multi_modal_projector") \
                 or name.startswith("vision_model") or name.startswith("audio_tower") \
-                or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector"):
+                or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector") \
+                or name.startswith("vision_tower."):
             # skip vision and audio tensors
             return
         yield from super().modify_tensors(data_torch, name, bid)
@@ -3815,14 +3823,14 @@ class DreamModel(TextModel):
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
 
-        vocab_dict = tokenizer.get_vocab()
+        vocab_dict = tokenizer.get_vocab()  # ty: ignore[unresolved-attribute]
         vocab_size = self.hparams.get("vocab_size", len(vocab_dict))
         assert max(vocab_dict.values()) < vocab_size
 
         tokpre = self.get_vocab_base_pre(tokenizer)
 
         reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab_dict.items()}
-        added_vocab = tokenizer.get_added_vocab()
+        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
@@ -3880,14 +3888,14 @@ class LLaDAModel(TextModel):
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
 
-        vocab_dict = tokenizer.get_vocab()
+        vocab_dict = tokenizer.get_vocab()  # ty: ignore[unresolved-attribute]
         vocab_size = self.hparams.get("vocab_size", len(vocab_dict))
         assert max(vocab_dict.values()) < vocab_size
 
         tokpre = self.get_vocab_base_pre(tokenizer)
 
         reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab_dict.items()}
-        added_vocab = tokenizer.get_added_vocab()
+        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
@@ -4665,9 +4673,9 @@ class Qwen3Model(Qwen2Model):
 
         self.is_rerank = True
         self.is_tied_embeddings = self.hparams.get("tie_word_embeddings", False)
-        self.token_false_id = tokenizer.convert_tokens_to_ids("no")
-        self.token_true_id = tokenizer.convert_tokens_to_ids("yes")
-        self.sep_token_id = tokenizer.convert_tokens_to_ids("|")
+        self.token_false_id = tokenizer.convert_tokens_to_ids("no")  # ty: ignore[unresolved-attribute, invalid-assignment]
+        self.token_true_id = tokenizer.convert_tokens_to_ids("yes")  # ty: ignore[unresolved-attribute, invalid-assignment]
+        self.sep_token_id = tokenizer.convert_tokens_to_ids("|")  # ty: ignore[unresolved-attribute]
 
         assert self.token_false_id is not None and self.token_true_id is not None
 
@@ -5936,7 +5944,7 @@ class KimiLinearModel(TextModel):
             # Build merges list using the approach similar to HunYuanMoE
             merges = []
             vocab = {}
-            mergeable_ranks = tokenizer.model._mergeable_ranks
+            mergeable_ranks = tokenizer.model._mergeable_ranks  # ty: ignore[unresolved-attribute]
             for token, rank in mergeable_ranks.items():
                 vocab[QwenModel.token_bytes_to_string(token)] = rank
                 if len(token) == 1:
@@ -5946,7 +5954,7 @@ class KimiLinearModel(TextModel):
                     merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
             # Build token list
             vocab_size = self.hparams["vocab_size"]
-            special_tokens = tokenizer.special_tokens
+            special_tokens = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]
             reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
             tokens: list[str] = []
             toktypes: list[int] = []
@@ -5972,7 +5980,7 @@ class KimiLinearModel(TextModel):
             special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
             special_vocab.add_to_gguf(self.gguf_writer)
             # override eos id in config.json with tiktoken eos id
-            self.gguf_writer.add_eos_token_id(tokenizer.eos_id)
+            self.gguf_writer.add_eos_token_id(tokenizer.eos_id)  # ty: ignore[unresolved-attribute]
         else:
             raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")
 
@@ -6466,11 +6474,11 @@ class BertModel(TextModel):
                 with open(tokenizer_config_path, "r", encoding="utf-8") as fp:
                     tokenizer_config_json = json.load(fp)
 
-            add_prefix = tokenizer.add_prefix_space
-            remove_whitespaces = tokenizer.clean_up_tokenization_spaces
+            add_prefix = tokenizer.add_prefix_space  # ty: ignore[unresolved-attribute]
+            remove_whitespaces = tokenizer.clean_up_tokenization_spaces  # ty: ignore[unresolved-attribute]
             precompiled_charsmap = b64decode(tokenizer_json["normalizer"]["precompiled_charsmap"])
 
-            vocab_size = max(self.hparams.get("vocab_size", 0), tokenizer.vocab_size)
+            vocab_size = max(self.hparams.get("vocab_size", 0), tokenizer.vocab_size)  # ty: ignore[unresolved-attribute]
         else:
             sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
             sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
@@ -6487,7 +6495,7 @@ class BertModel(TextModel):
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
-        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size  # ty: ignore[invalid-assignment]
 
         if isinstance(tokenizer, SentencePieceProcessor):
             for token_id in range(tokenizer.vocab_size()):
@@ -6509,20 +6517,20 @@ class BertModel(TextModel):
                 scores[token_id] = score
                 toktypes[token_id] = toktype
         else:
-            added_vocab = tokenizer.get_added_vocab()
+            added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
             unk_token = tokenizer_config_json.get("unk_token")
-            unk_token_id = added_vocab.get(unk_token, tokenizer_json["model"].get("unk_id", 3))
+            unk_token_id = added_vocab.get(unk_token, tokenizer_json["model"].get("unk_id", 3))  # ty: ignore[no-matching-overload]
 
-            for token_id in range(tokenizer.vocab_size):
-                piece = tokenizer._convert_id_to_token(token_id)
-                if (piece := tokenizer._convert_id_to_token(token_id)) is not None:
+            for token_id in range(tokenizer.vocab_size):  # ty: ignore[unresolved-attribute]
+                piece = tokenizer._convert_id_to_token(token_id)  # ty: ignore[unresolved-attribute]
+                if (piece := tokenizer._convert_id_to_token(token_id)) is not None:  # ty: ignore[unresolved-attribute]
                     text = piece.encode("utf-8")
                     score = tokenizer_json["model"]["vocab"][token_id][1]
 
                     toktype = SentencePieceTokenTypes.NORMAL
                     if token_id == unk_token_id:
                         toktype = SentencePieceTokenTypes.UNKNOWN
-                    elif token_id in tokenizer.all_special_ids:
+                    elif token_id in tokenizer.all_special_ids:  # ty: ignore[unresolved-attribute]
                         toktype = SentencePieceTokenTypes.CONTROL
                     elif token_id in added_vocab.values():
                         toktype = SentencePieceTokenTypes.USER_DEFINED
@@ -8831,7 +8839,7 @@ class DeepseekV2Model(TextModel):
             # Build merges list using the approach similar to HunYuanMoE
             merges = []
             vocab = {}
-            mergeable_ranks = tokenizer.model._mergeable_ranks
+            mergeable_ranks = tokenizer.model._mergeable_ranks  # ty: ignore[unresolved-attribute]
             for token, rank in mergeable_ranks.items():
                 vocab[QwenModel.token_bytes_to_string(token)] = rank
                 if len(token) == 1:
@@ -8842,7 +8850,7 @@ class DeepseekV2Model(TextModel):
 
             # Build token list
             vocab_size = self.hparams["vocab_size"]
-            special_tokens = tokenizer.special_tokens
+            special_tokens = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]
             reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
             tokens: list[str] = []
             toktypes: list[int] = []
@@ -9813,10 +9821,10 @@ class Glm4Model(TextModel):
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
         special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
-        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])
-        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
@@ -10044,12 +10052,12 @@ class ChatGLMModel(TextModel):
 
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
-        vocab_size = hparams.get("padded_vocab_size", len(tokenizer.get_vocab()))
-        assert max(tokenizer.get_vocab().values()) < vocab_size
+        vocab_size = hparams.get("padded_vocab_size", len(tokenizer.get_vocab()))  # ty: ignore[unresolved-attribute]
+        assert max(tokenizer.get_vocab().values()) < vocab_size  # ty: ignore[unresolved-attribute]
         role_special_tokens = ["<|system|>", "<|user|>", "<|assistant|>", "<|observation|>"]
         special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "sop", "eop"] + role_special_tokens
         for token_id in range(vocab_size):
-            piece = tokenizer._convert_id_to_token(token_id)
+            piece = tokenizer._convert_id_to_token(token_id)  # ty: ignore[unresolved-attribute]
             if token_id == 0:
                 piece = "<unk>"
             elif token_id == 1:
@@ -10057,17 +10065,17 @@ class ChatGLMModel(TextModel):
             elif token_id == 2:
                 piece = "<eos>"
 
-            text = piece.encode("utf-8")
+            text = piece.encode("utf-8")  # ty: ignore[unresolved-attribute]
             score = 0.0
             # Referencing the tokenizer Python implementation(https://huggingface.co/THUDM/chatglm3-6b/blob/main/tokenization_chatglm.py),
             # it is only valid if it is less than tokenizer.tokenizer.sp_model.vocab_size()
-            if len(piece) != 0 and token_id < tokenizer.tokenizer.sp_model.vocab_size():
-                score = tokenizer.tokenizer.sp_model.get_score(token_id)
+            if len(piece) != 0 and token_id < tokenizer.tokenizer.sp_model.vocab_size():  # ty: ignore[unresolved-attribute, invalid-argument-type]
+                score = tokenizer.tokenizer.sp_model.get_score(token_id)  # ty: ignore[unresolved-attribute]
 
-            if token_id >= tokenizer.tokenizer.sp_model.vocab_size():
+            if token_id >= tokenizer.tokenizer.sp_model.vocab_size():  # ty: ignore[unresolved-attribute]
                 if piece in special_tokens:
                     toktype = SentencePieceTokenTypes.CONTROL
-                elif len(piece) == 0:
+                elif len(piece) == 0:  # ty: ignore[invalid-argument-type]
                     text = f"[PAD{token_id}]".encode("utf-8")
                     toktype = SentencePieceTokenTypes.UNUSED
                 else:
@@ -10078,13 +10086,13 @@ class ChatGLMModel(TextModel):
                 continue
 
             toktype = SentencePieceTokenTypes.NORMAL
-            if tokenizer.tokenizer.sp_model.is_unknown(token_id):
+            if tokenizer.tokenizer.sp_model.is_unknown(token_id):  # ty: ignore[unresolved-attribute]
                 toktype = SentencePieceTokenTypes.UNKNOWN
-            elif tokenizer.tokenizer.sp_model.is_control(token_id):
+            elif tokenizer.tokenizer.sp_model.is_control(token_id):  # ty: ignore[unresolved-attribute]
                 toktype = SentencePieceTokenTypes.CONTROL
-            elif tokenizer.tokenizer.sp_model.is_unused(token_id):
+            elif tokenizer.tokenizer.sp_model.is_unused(token_id):  # ty: ignore[unresolved-attribute]
                 toktype = SentencePieceTokenTypes.UNUSED
-            elif tokenizer.tokenizer.sp_model.is_byte(token_id):
+            elif tokenizer.tokenizer.sp_model.is_byte(token_id):  # ty: ignore[unresolved-attribute]
                 toktype = SentencePieceTokenTypes.BYTE
 
             tokens.append(text)
@@ -10104,7 +10112,7 @@ class ChatGLMModel(TextModel):
 
     @staticmethod
     def token_bytes_to_string(b):
-        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
+        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode  # ty: ignore[unresolved-import]
         byte_encoder = bytes_to_unicode()
         return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])
 
@@ -10138,7 +10146,7 @@ class ChatGLMModel(TextModel):
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
         vocab_size = hparams.get("padded_vocab_size",hparams["vocab_size"])
-        assert max(tokenizer.get_vocab().values()) < vocab_size
+        assert max(tokenizer.get_vocab().values()) < vocab_size  # ty: ignore[unresolved-attribute]
 
         tokens, toktypes, tokpre = self.get_vocab_base()
         self.gguf_writer.add_tokenizer_model("gpt2")
@@ -10147,10 +10155,10 @@ class ChatGLMModel(TextModel):
         self.gguf_writer.add_token_types(toktypes)
         special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
         # only add special tokens when they were not already loaded from config.json
-        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
+        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # ty: ignore[unresolved-attribute]
         # this one is usually not in config.json anyway
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
@@ -11416,7 +11424,7 @@ class HunYuanMoEModel(TextModel):
         # 2. Reverse-engineer the merges list from mergeable_ranks
         merges = []
         vocab = {}
-        mergeable_ranks = tokenizer.mergeable_ranks
+        mergeable_ranks = tokenizer.mergeable_ranks  # ty: ignore[unresolved-attribute]
         for token, rank in mergeable_ranks.items():
             vocab[QwenModel.token_bytes_to_string(token)] = rank
             if len(token) == 1:
@@ -11427,8 +11435,8 @@ class HunYuanMoEModel(TextModel):
 
         # 3. Generate the tokens and toktypes lists
         vocab_size = self.hparams["vocab_size"]
-        assert tokenizer.vocab_size == vocab_size
-        special_tokens = tokenizer.special_tokens
+        assert tokenizer.vocab_size == vocab_size  # ty: ignore[unresolved-attribute]
+        special_tokens = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]
         reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
         tokens: list[str] = []
         toktypes: list[int] = []
@@ -11652,7 +11660,7 @@ class HunYuanModel(TextModel):
             # 2. Reverse-engineer the merges list from mergeable_ranks
             merges = []
             vocab = {}
-            mergeable_ranks = tokenizer.mergeable_ranks
+            mergeable_ranks = tokenizer.mergeable_ranks  # ty: ignore[unresolved-attribute]
             for token, rank in mergeable_ranks.items():
                 vocab[QwenModel.token_bytes_to_string(token)] = rank
                 if len(token) == 1:
@@ -11663,8 +11671,8 @@ class HunYuanModel(TextModel):
 
             # 3. Generate the tokens and toktypes lists
             vocab_size = self.hparams["vocab_size"]
-            assert tokenizer.vocab_size == vocab_size
-            special_tokens = tokenizer.special_tokens
+            assert tokenizer.vocab_size == vocab_size  # ty: ignore[unresolved-attribute]
+            special_tokens = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]
             reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
             tokens: list[str] = []
             toktypes: list[int] = []
@@ -12812,13 +12820,44 @@ class SolarOpenModel(Glm4MoeModel):
         self.gguf_writer.add_tokenizer_pre(tokpre)
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
-        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|endoftext|>"])
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<unk>"])
-        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|startoftext|>"])
+        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<unk>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|startoftext|>"])  # ty: ignore[unresolved-attribute]
         special_vocab.add_to_gguf(self.gguf_writer)
 
 
+@ModelBase.register("DotsOCRForCausalLM")
+class DotsOCRVisionModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        self.hparams_vision["image_size"] = 0 # dynamic resolution
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.DOTSOCR)
+        self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])
+        self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.find_vparam(["rms_norm_eps"]))
+        self.gguf_writer.add_vision_projector_scale_factor(self.find_vparam(["spatial_merge_size"]))
+        self.gguf_writer.add_vision_use_silu(True)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("vision_tower."):
+            if "vision_tower.blocks." in name and ".mlp." in name:
+                # note: to avoid naming conflicts in tensor_mapping.py, we need to handle FFN renaming here
+                # x = F.silu(self.fc1(x)) * self.fc3(x)
+                # x = self.fc2(x)
+                # fc1 -> gate, fc2 -> down, fc3 -> up
+                # mapping original names to Qwen2.5 naming scheme
+                name = name.replace("vision_tower.blocks.", "visual.blocks.")
+                name = name.replace(".fc1", ".gate_proj")
+                name = name.replace(".fc2", ".down_proj")
+                name = name.replace(".fc3", ".up_proj")
+            yield from super().modify_tensors(data_torch, name, bid)
+
+
 ###### CONVERSION LOGIC ######
 
 

convert_hf_to_gguf_update.py

@@ -296,7 +296,7 @@ for model in [*pre_computed_hashes, *all_models]:
         except Exception as e:
             raise OSError(f"Error loading tokenizer for model {name}.") from e
 
-        chktok = tokenizer.encode(CHK_TXT)
+        chktok = tokenizer.encode(CHK_TXT)  # ty: ignore[unresolved-attribute]
         chkhsh = sha256(str(chktok).encode()).hexdigest()
 
         logger.info(f"model: {name}")
@@ -468,7 +468,7 @@ for model in models:
 
     with open(f"models/ggml-vocab-{name}.gguf.out", "w") as f:
         for text in tests:
-            res = tokenizer.encode(text, add_special_tokens=False)
+            res = tokenizer.encode(text, add_special_tokens=False)  # ty: ignore[unresolved-attribute]
             for r in res:
                 f.write(f" {r}")
             f.write("\n")

convert_lora_to_gguf.py

@@ -402,7 +402,7 @@ if __name__ == '__main__':
                     # the invocation string includes the "<|start_of_turn|>"
                     # token, but the adapters themselves were trained to
                     # activate _after_ that first token, so we drop it here.
-                    alora_invocation_tokens = tokenizer(invocation_string)["input_ids"][1:]
+                    alora_invocation_tokens = tokenizer(invocation_string)["input_ids"][1:]  # ty: ignore[call-non-callable]
                 if alora_invocation_tokens:
                     logger.debug("GGUF KV: %s = %s", gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS, alora_invocation_tokens)
                     self.gguf_writer.add_key_value(

docs/multimodal.md

@@ -37,6 +37,7 @@ llama-server -hf ggml-org/gemma-3-4b-it-GGUF --no-mmproj-offload
 > - PaddleOCR-VL: https://github.com/ggml-org/llama.cpp/pull/18825
 > - GLM-OCR: https://github.com/ggml-org/llama.cpp/pull/19677
 > - Deepseek-OCR: https://github.com/ggml-org/llama.cpp/pull/17400
+> - Dots.OCR: https://github.com/ggml-org/llama.cpp/pull/17575
 > - HunyuanOCR: https://github.com/ggml-org/llama.cpp/pull/21395
 
 ## Pre-quantized models

examples/model-conversion/scripts/causal/run-casual-gen-embeddings-org.py

@@ -53,10 +53,10 @@ model_name = os.path.basename(model_path)
 print(f"Model name: {model_name}")
 
 prompt = "Hello world today"
-input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids  # ty: ignore[call-non-callable]
 print(f"Input tokens: {input_ids}")
 print(f"Input text: {repr(prompt)}")
-print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
+print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")  # ty: ignore[unresolved-attribute]
 
 with torch.no_grad():
     outputs = model(input_ids, output_hidden_states=True)
@@ -92,7 +92,7 @@ with torch.no_grad():
 
     # Print embeddings per token in the requested format
     print("\nToken embeddings:")
-    tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
+    tokens = tokenizer.convert_ids_to_tokens(input_ids[0])  # ty: ignore[unresolved-attribute]
     for i, embedding in enumerate(token_embeddings):
         # Format: show first few values, ..., then last few values
         if len(embedding) > 10:

examples/model-conversion/scripts/utils/semantic_check.py

@@ -207,8 +207,8 @@ def main():
         else:
             model = AutoModel.from_pretrained(args.model_path, trust_remote_code=True)
 
-    encoded = tokenizer(prompt, return_tensors="pt")
-    tokens = tokenizer.convert_ids_to_tokens(encoded['input_ids'][0])
+    encoded = tokenizer(prompt, return_tensors="pt")  # ty: ignore[call-non-callable]
+    tokens = tokenizer.convert_ids_to_tokens(encoded['input_ids'][0])  # ty: ignore[unresolved-attribute]
     n_tokens = len(tokens)
     print(f"n_tokens: {n_tokens}");
     print(f"hidden_size: {model.config.hidden_size}")

ggml/CMakeLists.txt

@@ -7,6 +7,8 @@ set(GGML_VERSION_MINOR 9)
 set(GGML_VERSION_PATCH 11)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
+list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
+
 find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
 if(GIT_EXE)
     # Get current git commit hash
@@ -204,12 +206,14 @@ option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"
 option(GGML_CUDA_FA                         "ggml: compile ggml FlashAttention CUDA kernels"  ON)
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
 option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})
+option(GGML_CUDA_NCCL                       "ggml: use NVIDIA Collective Comm. Library"       ON)
 set   (GGML_CUDA_COMPRESSION_MODE "size" CACHE STRING
                                             "ggml: cuda link binary compression mode; requires cuda 12.8+")
 set_property(CACHE GGML_CUDA_COMPRESSION_MODE PROPERTY STRINGS "none;speed;balance;size")
 
 option(GGML_HIP                             "ggml: use HIP"                                   OFF)
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
+option(GGML_HIP_RCCL                        "ggml: use ROCm Collective Comm. Library"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
 option(GGML_HIP_MMQ_MFMA                    "ggml: enable MFMA MMA for CDNA in MMQ"           ON)

ggml/cmake/FindNCCL.cmake

@@ -0,0 +1,36 @@
+# cmake/FindNCCL.cmake
+
+# NVIDIA does not distribute CMake files with NCCl, therefore use this file to find it instead.
+
+find_path(NCCL_INCLUDE_DIR
+    NAMES nccl.h
+    HINTS ${NCCL_ROOT} $ENV{NCCL_ROOT} $ENV{CUDA_HOME} /usr/local/cuda
+    PATH_SUFFIXES include
+)
+
+find_library(NCCL_LIBRARY
+    NAMES nccl
+    HINTS ${NCCL_ROOT} $ENV{NCCL_ROOT} $ENV{CUDA_HOME} /usr/local/cuda
+    PATH_SUFFIXES lib lib64
+)
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(NCCL
+    DEFAULT_MSG
+    NCCL_LIBRARY NCCL_INCLUDE_DIR
+)
+
+if(NCCL_FOUND)
+    set(NCCL_LIBRARIES ${NCCL_LIBRARY})
+    set(NCCL_INCLUDE_DIRS ${NCCL_INCLUDE_DIR})
+
+    if(NOT TARGET NCCL::NCCL)
+        add_library(NCCL::NCCL UNKNOWN IMPORTED)
+        set_target_properties(NCCL::NCCL PROPERTIES
+            IMPORTED_LOCATION "${NCCL_LIBRARY}"
+            INTERFACE_INCLUDE_DIRECTORIES "${NCCL_INCLUDE_DIR}"
+        )
+    endif()
+endif()
+
+mark_as_advanced(NCCL_INCLUDE_DIR NCCL_LIBRARY)

ggml/include/ggml-backend.h

@@ -68,7 +68,7 @@ extern "C" {
     GGML_API void                           ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
 
     // tensor copy between different backends
-    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API void ggml_backend_tensor_copy(const struct ggml_tensor * src, struct ggml_tensor * dst);
 
     //
     // Backend (stream)
@@ -83,13 +83,17 @@ extern "C" {
     GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
     GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
 
-    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_set_async   (ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async   (ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_set_2d_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+    GGML_API void ggml_backend_tensor_get_2d_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
 
     // "offset" refers to the offset in tensor->data for setting/getting data
-    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_memset(   struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_set   (      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get   (const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_set_2d(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+    GGML_API void ggml_backend_tensor_get_2d(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+    GGML_API void ggml_backend_tensor_memset(      struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
 
     GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
 
@@ -109,7 +113,7 @@ extern "C" {
     // the copy is performed after all the currently queued operations in backend_src
     // backend_dst will wait for the copy to complete before performing other operations
     // automatic fallback to sync copy if async is not supported
-    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
     GGML_API ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend);
 
@@ -135,7 +139,9 @@ extern "C" {
         // integrated GPU device using host memory
         GGML_BACKEND_DEVICE_TYPE_IGPU,
         // accelerator devices intended to be used together with the CPU backend (e.g. BLAS or AMX)
-        GGML_BACKEND_DEVICE_TYPE_ACCEL
+        GGML_BACKEND_DEVICE_TYPE_ACCEL,
+        // "meta" device wrapping multiple other devices for tensor parallelism
+        GGML_BACKEND_DEVICE_TYPE_META,
     };
 
     // functionality supported by the device
@@ -196,7 +202,9 @@ extern "C" {
 
     // Common functions that may be obtained using ggml_backend_reg_get_proc_address
 
-    // Split buffer type for tensor parallelism
+    // AllReduce operation for tensor parallelism (meta backend)
+    typedef bool                         (*ggml_backend_allreduce_tensor_t)(ggml_backend_t * backends, struct ggml_tensor ** tensors, size_t n_backends);
+    // Split buffer type for tensor parallelism (old)
     typedef ggml_backend_buffer_type_t   (*ggml_backend_split_buffer_type_t)(int main_device, const float * tensor_split);
     // Set the number of threads for the backend
     typedef void                         (*ggml_backend_set_n_threads_t)(ggml_backend_t backend, int n_threads);

ggml/include/ggml-cuda.h

@@ -27,6 +27,9 @@ GGML_BACKEND_API bool ggml_backend_is_cuda(ggml_backend_t backend);
 // device buffer
 GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
 
+// conduct allreduce operation between devices
+GGML_BACKEND_API bool ggml_backend_cuda_allreduce_tensor(ggml_backend_t * backends, struct ggml_tensor ** tensors, size_t n_backends);
+
 // split tensor buffer that splits matrices by rows across multiple devices
 GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
 

ggml/src/CMakeLists.txt

@@ -200,6 +200,7 @@ add_library(ggml-base
             ggml.cpp
             ggml-alloc.c
             ggml-backend.cpp
+            ggml-backend-meta.cpp
             ggml-opt.cpp
             ggml-threading.cpp
             ggml-threading.h

ggml/src/ggml-alloc.c

@@ -1236,6 +1236,9 @@ size_t ggml_backend_alloc_ctx_tensors_from_buft_size(struct ggml_context * ctx,
 
 ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
     size_t nbytes_total = 0;
+    if (ggml_backend_buft_is_meta(buft)) {
+        return ggml_backend_meta_alloc_ctx_tensors_from_buft(ctx, buft);
+    }
     return ggml_backend_alloc_ctx_tensors_from_buft_impl(ctx, buft, &nbytes_total, /*no_alloc =*/ false);
 }
 

ggml/src/ggml-backend-impl.h

@@ -49,6 +49,10 @@ extern "C" {
         void         (*memset_tensor)(ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
         void         (*set_tensor)   (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void         (*get_tensor)   (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        // (optional) 2d data copies
+        void         (*set_tensor_2d)(ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+        void         (*get_tensor_2d)(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+
         // (optional) tensor copy: dst is in the buffer, src may be in any buffer, including buffers from a different backend (return false if not supported)
         bool         (*cpy_tensor)   (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst);
         // clear the entire buffer
@@ -80,6 +84,20 @@ extern "C" {
     GGML_API bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
     GGML_API void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
 
+    //
+    // Backend (meta)
+    //
+
+    GGML_API bool ggml_backend_is_meta       (ggml_backend_t backend);
+    GGML_API bool ggml_backend_buffer_is_meta(ggml_backend_buffer_t buf);
+    GGML_API bool ggml_backend_buft_is_meta  (ggml_backend_buffer_type_t buft);
+
+    GGML_API size_t         ggml_backend_meta_n_backends    (ggml_backend_t meta_backend);
+    GGML_API ggml_backend_t ggml_backend_meta_simple_backend(ggml_backend_t meta_backend, size_t index);
+
+    // temporary workaround to statically allocate tensors from a context in a deduplicated way:
+    GGML_API struct ggml_backend_buffer * ggml_backend_meta_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
+
     //
     // Backend (stream)
     //
@@ -90,8 +108,10 @@ extern "C" {
         void (*free)(ggml_backend_t backend);
 
         // (optional) asynchronous tensor data access
-        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        void (*set_tensor_async)   (ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*get_tensor_async)   (ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        void (*set_tensor_2d_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+        void (*get_tensor_2d_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
         bool (*cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
         // (optional) complete all pending operations (required if the backend supports async operations)

ggml/src/ggml-backend.cpp

@@ -123,7 +123,7 @@ size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
 void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
     GGML_ASSERT(buffer);
     // get_base is optional if the buffer is zero-sized
-    if (buffer->size == 0) {
+    if (!ggml_backend_buffer_is_meta(buffer) && buffer->size == 0) {
         return NULL;
     }
 
@@ -279,15 +279,57 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
     }
 }
 
+void ggml_backend_tensor_set_2d_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size,
+            size_t n_copies, size_t stride_tensor, size_t stride_data) {
+    GGML_ASSERT(backend);
+    GGML_ASSERT(tensor);
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
+    if (n_copies <= 1 || backend->iface.set_tensor_2d_async == NULL) {
+        for (size_t i = 0; i < n_copies; i++) {
+            ggml_backend_tensor_set_async(backend, tensor, (const char *) data + i*stride_data, offset + i*stride_tensor, size);
+        }
+        return;
+    }
+    if (size == 0) {
+        return;
+    }
+
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + (n_copies-1)*stride_tensor + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+    backend->iface.set_tensor_2d_async(backend, tensor, data, offset, size, n_copies, stride_tensor, stride_data);
+}
+
+void ggml_backend_tensor_get_2d_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size,
+            size_t n_copies, size_t stride_tensor, size_t stride_data) {
+    GGML_ASSERT(backend);
+    GGML_ASSERT(tensor);
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
+    if (n_copies <= 1 || backend->iface.set_tensor_2d_async == NULL) {
+        for (size_t i = 0; i < n_copies; i++) {
+            ggml_backend_tensor_get_async(backend, tensor, (char *) data + i*stride_data, offset + i*stride_tensor, size);
+        }
+        return;
+    }
+    if (size == 0) {
+        return;
+    }
+
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + (n_copies-1)*stride_tensor + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+    backend->iface.get_tensor_2d_async(backend, tensor, data, offset, size, n_copies, stride_tensor, stride_data);
+}
+
 void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     GGML_ASSERT(tensor);
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
 
     if (size == 0) {
         return;
     }
 
-    GGML_ASSERT(buf != NULL && "tensor buffer not set");
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
@@ -297,18 +339,62 @@ void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, siz
 void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     GGML_ASSERT(tensor);
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
 
     if (size == 0) {
         return;
     }
 
-    GGML_ASSERT(buf != NULL && "tensor buffer not set");
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
     buf->iface.get_tensor(buf, tensor, data, offset, size);
 }
 
+void ggml_backend_tensor_set_2d(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size,
+            size_t n_copies, size_t stride_tensor, size_t stride_data) {
+    GGML_ASSERT(tensor);
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+
+    if (n_copies <= 1 || buf->iface.set_tensor_2d == NULL) {
+        for (size_t i = 0; i < n_copies; i++) {
+            ggml_backend_tensor_set(tensor, (const char *) data + i*stride_data, offset + i*stride_tensor, size);
+        }
+        return;
+    }
+    if (size == 0) {
+        return;
+    }
+
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + (n_copies-1)*stride_tensor + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+
+    buf->iface.set_tensor_2d(buf, tensor, data, offset, size, n_copies, stride_tensor, stride_data);
+}
+
+void ggml_backend_tensor_get_2d(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size,
+            size_t n_copies, size_t stride_tensor, size_t stride_data) {
+    GGML_ASSERT(tensor);
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+
+    if (n_copies <= 1 || buf->iface.set_tensor_2d == NULL) {
+        for (size_t i = 0; i < n_copies; i++) {
+            ggml_backend_tensor_get(tensor, (char *) data + i*stride_data, offset + i*stride_tensor, size);
+        }
+        return;
+    }
+    if (size == 0) {
+        return;
+    }
+
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + (n_copies-1)*stride_tensor + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+
+    buf->iface.get_tensor_2d(buf, tensor, data, offset, size, n_copies, stride_tensor, stride_data);
+}
+
 void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
     GGML_ASSERT(tensor);
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -388,7 +474,7 @@ ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend) {
 
 // backend copy
 
-void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
+void ggml_backend_tensor_copy(const struct ggml_tensor * src, struct ggml_tensor * dst) {
     GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
 
     if (src == dst) {
@@ -402,7 +488,7 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
     } else if (!ggml_backend_buffer_copy_tensor(src, dst)) {
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
-#endif
+#endif // NDEBUG
         size_t nbytes = ggml_nbytes(src);
         void * data = malloc(nbytes);
         ggml_backend_tensor_get(src, data, 0, nbytes);
@@ -411,7 +497,7 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
     }
 }
 
-void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst) {
+void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst) {
     GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
 
     if (src == dst) {
@@ -500,6 +586,7 @@ enum ggml_backend_dev_type ggml_backend_dev_type(ggml_backend_dev_t device) {
 }
 
 void ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props) {
+    GGML_ASSERT(device);
     memset(props, 0, sizeof(*props));
     device->iface.get_props(device, props);
 }
@@ -610,6 +697,8 @@ static const struct ggml_backend_buffer_i ggml_backend_multi_buffer_i = {
     /* .memset_tensor   = */ NULL,
     /* .set_tensor      = */ NULL,
     /* .get_tensor      = */ NULL,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ NULL,
     /* .clear           = */ ggml_backend_multi_buffer_clear,
     /* .reset           = */ NULL,
@@ -1899,8 +1988,9 @@ enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct
     GGML_ASSERT(tensor->data == NULL);
     GGML_ASSERT(tensor->view_src == NULL);
     GGML_ASSERT(addr >= ggml_backend_buffer_get_base(buffer));
-    GGML_ASSERT((char *)addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <=
-                (char *)ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer));
+    GGML_ASSERT(ggml_backend_buffer_is_meta(buffer) ||
+        (char *) addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <=
+        (char *) ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer));
 
     tensor->buffer = buffer;
     tensor->data = addr;
@@ -2174,6 +2264,8 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
     /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_cpu_buffer_clear,
     /* .reset           = */ NULL,
@@ -2186,6 +2278,8 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
     /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_cpu_buffer_clear,
     /* .reset           = */ NULL,

ggml/src/ggml-blas/ggml-blas.cpp

@@ -262,6 +262,8 @@ static struct ggml_backend_i blas_backend_i = {
     /* .get_name                = */ ggml_backend_blas_get_name,
     /* .free                    = */ ggml_backend_blas_free,
     /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
     /* .synchronize             = */ NULL,

ggml/src/ggml-cann/ggml-cann.cpp

@@ -1457,6 +1457,8 @@ static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
     /* .memset_tensor   = */ NULL,
     /* .set_tensor      = */ ggml_backend_cann_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_cann_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ ggml_backend_cann_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_cann_buffer_clear,
     /* .reset           = */ NULL,
@@ -2698,6 +2700,8 @@ static const ggml_backend_i ggml_backend_cann_interface = {
     /* .free                    = */ ggml_backend_cann_free,
     /* .set_tensor_async        = */ ggml_backend_cann_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_cann_get_tensor_async,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ ggml_backend_cann_cpy_tensor_async,
     /* .synchronize             = */ ggml_backend_cann_synchronize,
     /* .graph_plan_create       = */ NULL,

ggml/src/ggml-cpu/amx/amx.cpp

@@ -111,6 +111,8 @@ static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
     /* .memset_tensor   = */ ggml_backend_amx_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_amx_buffer_set_tensor,
     /* .get_tensor      = */ nullptr,
+    /* .set_tensor_2d   = */ nullptr,
+    /* .get_tensor_2d   = */ nullptr,
     /* .cpy_tensor      = */ nullptr,
     /* .clear           = */ ggml_backend_amx_buffer_clear,
     /* .reset           = */ nullptr,

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -195,6 +195,8 @@ static const struct ggml_backend_i ggml_backend_cpu_i = {
     /* .free                    = */ ggml_backend_cpu_free,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
     /* .synchronize             = */ NULL,
     /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,

ggml/src/ggml-cuda/CMakeLists.txt

@@ -181,6 +181,16 @@ if (CUDAToolkit_FOUND)
         target_link_libraries(ggml-cuda PRIVATE CUDA::cuda_driver)
     endif()
 
+    if (GGML_CUDA_NCCL)
+        find_package(NCCL)
+        if (NCCL_FOUND)
+            add_compile_definitions(GGML_USE_NCCL)
+            target_link_libraries(ggml-cuda PRIVATE NCCL::NCCL)
+        else()
+            message(STATUS "Warning: NCCL not found, performance for multiple CUDA GPUs will be suboptimal")
+        endif()
+    endif()
+
     set(CUDA_CXX_FLAGS "")
 
     set(CUDA_FLAGS -use_fast_math -extended-lambda)

ggml/src/ggml-cuda/argsort.cu

@@ -60,24 +60,24 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
 
     if (order == GGML_SORT_ORDER_ASC) {
         if (nrows == 1) {
-            DeviceRadixSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
+            CUDA_CHECK(DeviceRadixSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
                                        temp_indices, dst,                                  // values (indices)
-                                       ncols, 0, sizeof(float) * 8, stream);
+                                       ncols, 0, sizeof(float) * 8, stream));
         } else {
-            DeviceSegmentedSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
+            CUDA_CHECK(DeviceSegmentedSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
                                            temp_indices, dst,                                  // values (indices)
                                            ncols * nrows, nrows,  // num items, num segments
-                                           offset_iterator, offset_iterator + 1, stream);
+                                           offset_iterator, offset_iterator + 1, stream));
         }
     } else {
         if (nrows == 1) {
-            DeviceRadixSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
+            CUDA_CHECK(DeviceRadixSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
                                                  temp_indices, dst,                                  // values (indices)
-                                                 ncols, 0, sizeof(float) * 8, stream);
+                                                 ncols, 0, sizeof(float) * 8, stream));
         } else {
-            DeviceSegmentedSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys, temp_indices,
+            CUDA_CHECK(DeviceSegmentedSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys, temp_indices,
                                                      dst, ncols * nrows, nrows, offset_iterator, offset_iterator + 1,
-                                                     stream);
+                                                     stream));
         }
     }
 
@@ -86,22 +86,22 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
 
     if (order == GGML_SORT_ORDER_ASC) {
         if (nrows == 1) {
-            DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
+            CUDA_CHECK(DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
                                        temp_indices, dst,  // values (indices)
-                                       ncols, 0, sizeof(float) * 8, stream);
+                                       ncols, 0, sizeof(float) * 8, stream));
         } else {
-            DeviceSegmentedSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst,
-                                           ncols * nrows, nrows, offset_iterator, offset_iterator + 1, stream);
+            CUDA_CHECK(DeviceSegmentedSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst,
+                                           ncols * nrows, nrows, offset_iterator, offset_iterator + 1, stream));
         }
     } else {
         if (nrows == 1) {
-            DeviceRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
+            CUDA_CHECK(DeviceRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
                                                  temp_indices, dst,                                  // values (indices)
-                                                 ncols, 0, sizeof(float) * 8, stream);
+                                                 ncols, 0, sizeof(float) * 8, stream));
         } else {
-            DeviceSegmentedSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
+            CUDA_CHECK(DeviceSegmentedSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
                                                      temp_indices, dst, ncols * nrows, nrows, offset_iterator,
-                                                     offset_iterator + 1, stream);
+                                                     offset_iterator + 1, stream));
         }
     }
 }

ggml/src/ggml-cuda/binbcast.cu

@@ -472,6 +472,36 @@ void ggml_cuda_op_fused_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst,
     }
 }
 
+void ggml_cuda_op_fused_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst, int n_fuse) {
+    GGML_ASSERT(2 <= n_fuse && n_fuse <= 8);
+
+    switch (n_fuse) {
+        case 2:
+            ggml_cuda_op_fused_binbcast_impl<op_mul, 2>(ctx, dst);
+            break;
+        case 3:
+            ggml_cuda_op_fused_binbcast_impl<op_mul, 3>(ctx, dst);
+            break;
+        case 4:
+            ggml_cuda_op_fused_binbcast_impl<op_mul, 4>(ctx, dst);
+            break;
+        case 5:
+            ggml_cuda_op_fused_binbcast_impl<op_mul, 5>(ctx, dst);
+            break;
+        case 6:
+            ggml_cuda_op_fused_binbcast_impl<op_mul, 6>(ctx, dst);
+            break;
+        case 7:
+            ggml_cuda_op_fused_binbcast_impl<op_mul, 7>(ctx, dst);
+            break;
+        case 8:
+            ggml_cuda_op_fused_binbcast_impl<op_mul, 8>(ctx, dst);
+            break;
+        default:
+            GGML_ASSERT(false && "Unsupported n_fuse value");
+    }
+}
+
 void ggml_cuda_op_repeat_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
 

ggml/src/ggml-cuda/binbcast.cuh

@@ -9,3 +9,4 @@ void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 void ggml_cuda_op_repeat_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_fused_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst, int n_fuse);
+void ggml_cuda_op_fused_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst, int n_fuse);

ggml/src/ggml-cuda/common.cuh

@@ -67,6 +67,7 @@
 #define GGML_CUDA_CC_CDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x908)  // MI100, minimum for MFMA, acc registers
 #define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x90a)  // MI210 (gfx90a), minimum acc register renaming
 #define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x942)  // MI300
+#define GGML_CUDA_CC_CDNA4      (GGML_CUDA_CC_OFFSET_AMD + 0x950)  // MI350X/MI355X
 
 // RDNA removes MFMA, dp4a, xnack, acc registers, wave size is 32
 #define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
@@ -87,7 +88,8 @@
 #define GGML_CUDA_CC_IS_CDNA(cc)    (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)
 #define GGML_CUDA_CC_IS_CDNA1(cc)   (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2)
 #define GGML_CUDA_CC_IS_CDNA2(cc)   (cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)
-#define GGML_CUDA_CC_IS_CDNA3(cc)   (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_CDNA3(cc)   (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_CDNA4)
+#define GGML_CUDA_CC_IS_CDNA4(cc)   (cc >= GGML_CUDA_CC_CDNA4 && cc < GGML_CUDA_CC_RDNA1)
 
 // Moore Threads
 #define MUSART_HMASK 40300 // MUSA rc4.3, min. ver. for half2 -> uint mask comparisons
@@ -186,6 +188,10 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
 
 #define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
 
+#ifdef GGML_USE_NCCL
+#define NCCL_CHECK(err) CUDA_CHECK_GEN(err, ncclSuccess, ncclGetErrorString)
+#endif // GGML_USE_NCCL
+
 #if !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
 static const char * cu_get_error_str(CUresult err) {
     const char * err_str;
@@ -1086,6 +1092,10 @@ struct ggml_cuda_device_info {
     cuda_device_info devices[GGML_CUDA_MAX_DEVICES] = {};
 
     std::array<float, GGML_CUDA_MAX_DEVICES> default_tensor_split = {};
+
+#ifdef GGML_USE_NCCL
+    ncclComm_t comms[GGML_CUDA_MAX_DEVICES];
+#endif // GGML_USE_NCCL
 };
 
 const ggml_cuda_device_info & ggml_cuda_info();

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -324,6 +324,28 @@ static ggml_cuda_device_info ggml_cuda_init() {
     // configure logging to stdout
     // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
 
+    for (int id = 0; id < info.device_count; ++id) {
+        ggml_cuda_set_device(id);
+        for (int id_other = 0; id_other < info.device_count; ++id_other) {
+            if (id == id_other) {
+                continue;
+            }
+            int can_access_peer;
+            CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
+            if (can_access_peer) {
+                CUDA_CHECK(cudaDeviceEnablePeerAccess(id_other, 0));
+            }
+        }
+    }
+
+#ifdef GGML_USE_NCCL
+    int dev_ids[GGML_CUDA_MAX_DEVICES];
+    for (int id = 0; id < info.device_count; ++id) {
+        dev_ids[id] = id;
+    }
+    NCCL_CHECK(ncclCommInitAll(info.comms, info.device_count, dev_ids));
+#endif // GGML_USE_NCCL
+
     return info;
 }
 
@@ -632,26 +654,46 @@ static enum ggml_status ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer
 }
 
 static void ggml_backend_cuda_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
-    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *) buffer->context;
 
     ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + offset, value, size, cudaStreamPerThread));
+    CUDA_CHECK(cudaMemsetAsync((char *) tensor->data + offset, value, size, cudaStreamPerThread));
     CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
 static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *) buffer->context;
 
     ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cudaStreamPerThread));
+    CUDA_CHECK(cudaMemcpyAsync((char *) tensor->data + offset, data, size, cudaMemcpyHostToDevice, cudaStreamPerThread));
     CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
 static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *) buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaMemcpyAsync(data, (const char *) tensor->data + offset, size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+}
+
+static void ggml_backend_cuda_buffer_set_tensor_2d(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data,
+        size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *) buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaMemcpy2DAsync(
+        (char *) tensor->data + offset, stride_tensor, data, stride_data, size, n_copies, cudaMemcpyHostToDevice, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+}
+
+static void ggml_backend_cuda_buffer_get_tensor_2d(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data,
+        size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
+    CUDA_CHECK(cudaMemcpy2DAsync(
+        data, stride_data, (const char *) tensor->data + offset, stride_tensor, size, n_copies, cudaMemcpyDeviceToHost, cudaStreamPerThread));
     CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
@@ -691,6 +733,8 @@ static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
     /* .memset_tensor   = */ ggml_backend_cuda_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_cuda_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_cuda_buffer_get_tensor,
+    /* .set_tensor_2d   = */ ggml_backend_cuda_buffer_set_tensor_2d,
+    /* .get_tensor_2d   = */ ggml_backend_cuda_buffer_get_tensor_2d,
     /* .cpy_tensor      = */ ggml_backend_cuda_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_cuda_buffer_clear,
     /* .reset           = */ NULL,
@@ -1003,6 +1047,8 @@ static const ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = {
     /* .memset_tensor   = */ NULL,
     /* .set_tensor      = */ ggml_backend_cuda_split_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_cuda_split_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ NULL,
     /* .clear           = */ ggml_backend_cuda_split_buffer_clear,
     /* .reset           = */ NULL,
@@ -1079,6 +1125,83 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_inte
     /* .is_host          = */ ggml_backend_cuda_split_buffer_type_is_host,
 };
 
+bool ggml_backend_cuda_allreduce_tensor(ggml_backend_t * backends, struct ggml_tensor ** tensors, size_t n_backends) {
+#ifdef GGML_USE_NCCL
+    const int64_t ne = ggml_nelements(tensors[0]);
+    // FIXME the input of llm_graph_context::build_in_out_ids can produce a tensor with 0 elements if n_outputs == 0
+    // This then causes a crash in this function
+    if (ne == 0) {
+        return true;
+    }
+    for (size_t i = 0; i < n_backends; ++i) {
+        GGML_ASSERT(tensors[i] != nullptr);
+        GGML_ASSERT(ggml_nelements(tensors[i]) == ne);
+        GGML_ASSERT(ggml_is_contiguously_allocated(tensors[i]));
+    }
+
+    const ggml_cuda_device_info info = ggml_cuda_info();
+
+    // For small tensors, simply reduce them as FP32.
+    // The following heuristic for how "small" a tensor should be is based on RTX 4090s connected via 16x PCIe 4.0.
+    if ((n_backends <= 2 && ne < 32768) || (n_backends == 3 && ne < 131072) || (n_backends >= 4 && ne < 262144)) {
+        NCCL_CHECK(ncclGroupStart());
+        for (size_t i = 0; i < n_backends; ++i) {
+            ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
+            NCCL_CHECK(ncclAllReduce(tensors[i]->data, tensors[i]->data, ne, ncclFloat, ncclSum, info.comms[cuda_ctx->device], cuda_ctx->stream()));
+        }
+        NCCL_CHECK(ncclGroupEnd());
+
+        return true;
+    }
+
+    // For large tensors it's faster to compress them to BF16 for the reduction:
+    to_bf16_cuda_t to_bf16 = ggml_get_to_bf16_cuda(GGML_TYPE_F32);
+    to_fp32_cuda_t to_fp32 = ggml_get_to_fp32_cuda(GGML_TYPE_BF16);
+
+    ggml_cuda_pool_alloc<nv_bfloat16> tmp[GGML_CUDA_MAX_DEVICES];
+    for (size_t i = 0; i < n_backends; ++i) {
+        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
+        tmp[i].pool = &cuda_ctx->pool();
+        tmp[i].alloc(ne);
+
+        ggml_cuda_set_device(i);
+        to_bf16(tensors[i]->data, tmp[i].get(), ne, cuda_ctx->stream());
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    NCCL_CHECK(ncclGroupStart());
+    for (size_t i = 0; i < n_backends; ++i) {
+        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
+        NCCL_CHECK(ncclAllReduce(tmp[i].get(), tmp[i].get(), ne, ncclBfloat16, ncclSum, info.comms[cuda_ctx->device], cuda_ctx->stream()));
+    }
+    NCCL_CHECK(ncclGroupEnd());
+
+    for (size_t i = 0; i < n_backends; ++i) {
+        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
+
+        ggml_cuda_set_device(i);
+        to_fp32(tmp[i].get(), (float *) tensors[i]->data, ne, cuda_ctx->stream());
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    return true;
+#else
+    // If NCCL is installed it is used by default for optimal performance.
+    // However, NVIDIA does not distribute NCCL with CUDA so users may be unwittingly missing this package.
+    // RCCL is disabled by default, users are explicitly opting in.
+    // Therefore print no warning for RCCL.
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    static bool warning_printed = false;
+    if (!warning_printed) {
+        GGML_LOG_WARN("%s: NVIDIA Collective Communications Library (NCCL) is unavailable, multi GPU performance will be suboptimal\n", __func__);
+        warning_printed = true;
+    }
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    GGML_UNUSED_VARS(backends, tensors, n_backends);
+    return false;
+#endif // GGML_USE_NCCL
+}
+
 ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
@@ -1425,64 +1548,6 @@ static void ggml_cuda_op_mul_mat_cublas(
     GGML_UNUSED_VARS(dst, src1_ddq_i, src1_padded_row_size);
 }
 
-static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
-    static bool peer_access_enabled = false;
-
-    const bool enable_peer_access = n_tokens <= GGML_CUDA_PEER_MAX_BATCH_SIZE;
-
-    if (peer_access_enabled == enable_peer_access) {
-        return;
-    }
-
-#ifdef NDEBUG
-    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
-        ggml_cuda_set_device(id);
-        CUDA_CHECK(cudaDeviceSynchronize());
-    }
-
-    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
-        ggml_cuda_set_device(id);
-
-        for (int id_other = 0; id_other < ggml_backend_cuda_get_device_count(); ++id_other) {
-            if (id == id_other) {
-                continue;
-            }
-            if (id != main_device && id_other != main_device) {
-                continue;
-            }
-
-            int can_access_peer;
-            CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
-            if (can_access_peer) {
-                if (enable_peer_access) {
-                    cudaError_t err = cudaDeviceEnablePeerAccess(id_other, 0);
-                    if (err != cudaErrorPeerAccessAlreadyEnabled) {
-                        CUDA_CHECK(err);
-                    } else {
-                        // reset the error
-                        (void)cudaGetLastError();
-                    }
-                } else {
-                    cudaError_t err = cudaDeviceDisablePeerAccess(id_other);
-                    if (err != cudaErrorPeerAccessNotEnabled) {
-                        CUDA_CHECK(err);
-                    } else {
-                        // reset the error
-                        (void)cudaGetLastError();
-                    }
-                }
-            }
-        }
-    }
-
-    ggml_cuda_set_device(main_device);
-#endif // NDEBUG
-
-    peer_access_enabled = enable_peer_access;
-
-    GGML_UNUSED(main_device);
-}
-
 static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
     void * dst, int dstDevice, size_t dpitch, void * src, int srcDevice, size_t spitch, size_t width, size_t height, cudaStream_t stream) {
 
@@ -2483,11 +2548,6 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
 }
 
 static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) {
-    // why is this here instead of mul_mat?
-    if (dst->src[0] != nullptr && ggml_backend_buft_is_cuda_split(dst->src[0]->buffer->buft)) {
-        ggml_cuda_set_peer_access(dst->src[1]->ne[1], ctx.device);
-    }
-
     switch (dst->op) {
         case GGML_OP_ARGMAX:
             ggml_cuda_argmax(ctx, dst);
@@ -2845,21 +2905,43 @@ static void ggml_backend_cuda_free(ggml_backend_t backend) {
 }
 
 static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
 
-    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cuda_ctx->stream()));
+    CUDA_CHECK(cudaMemcpyAsync((char *) tensor->data + offset, data, size, cudaMemcpyHostToDevice, cuda_ctx->stream()));
 }
 
 static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
 
-    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cuda_ctx->stream()));
+    CUDA_CHECK(cudaMemcpyAsync(data, (const char *) tensor->data + offset, size, cudaMemcpyDeviceToHost, cuda_ctx->stream()));
+}
+
+static void ggml_backend_cuda_set_tensor_2d_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data,
+        size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
+
+    CUDA_CHECK(cudaMemcpy2DAsync(
+        (char *) tensor->data + offset, stride_tensor, data, stride_data, size, n_copies, cudaMemcpyHostToDevice, cuda_ctx->stream()));
+}
+
+static void ggml_backend_cuda_get_tensor_2d_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data,
+        size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
+
+    CUDA_CHECK(cudaMemcpy2DAsync(
+        data, stride_data, (const char *) tensor->data + offset, stride_tensor, size, n_copies, cudaMemcpyDeviceToHost, cuda_ctx->stream()));
 }
 
 static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
@@ -2870,21 +2952,21 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
         return false;
     }
 
-    if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
+    if (!ggml_backend_buffer_is_cuda(buf_src) || !ggml_backend_buffer_is_cuda(buf_dst)) {
         return false;
     }
 
     // device -> device copy
-    ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
-    ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
+    ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *) backend_src->context;
+    ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *) backend_dst->context;
 
-    ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
-    ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
+    ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *) buf_src->context;
+    ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *) buf_dst->context;
 
     if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: backend and buffer devices do not match\n", __func__);
-#endif
+#endif // NDEBUG
         return false;
     }
 
@@ -2897,7 +2979,7 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
             return false;
 #else
             CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, cuda_ctx_dst->device, src->data, cuda_ctx_src->device, ggml_nbytes(dst), cuda_ctx_src->stream()));
-#endif
+#endif // GGML_CUDA_NO_PEER_COPY
         }
 
         // record event on src stream after the copy
@@ -3676,10 +3758,10 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                         continue;
                     }
 
-                    if (node->op == GGML_OP_ADD) {
+                    if (node->op == GGML_OP_ADD || node->op == GGML_OP_MUL) {
                         int n_fuse = 0;
                         ggml_op ops[8];
-                        std::fill(ops, ops + 8, GGML_OP_ADD);
+                        std::fill(ops, ops + 8, node->op);
 
                         for (; n_fuse <= 6; ++n_fuse){
                             if (!ggml_can_fuse(cgraph, i + n_fuse, ops + n_fuse, 2)) {
@@ -3696,13 +3778,17 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                         n_fuse++;
 
                         if (n_fuse > 1) {
-                            ggml_tensor fused_add_node;
-                            memcpy(&fused_add_node, node, sizeof(ggml_tensor));
+                            ggml_tensor fused_node;
+                            memcpy(&fused_node, node, sizeof(ggml_tensor));
                             for (int j = 0; j < n_fuse - 1; ++j) {
-                                fused_add_node.src[j + 2] = cgraph->nodes[i + j + 1]->src[1];
+                                fused_node.src[j + 2] = cgraph->nodes[i + j + 1]->src[1];
+                            }
+                            fused_node.data = cgraph->nodes[i + n_fuse - 1]->data;
+                            if (node->op == GGML_OP_ADD) {
+                                ggml_cuda_op_fused_add(*cuda_ctx, &fused_node, n_fuse);
+                            } else {
+                                ggml_cuda_op_fused_mul(*cuda_ctx, &fused_node, n_fuse);
                             }
-                            fused_add_node.data = cgraph->nodes[i + n_fuse - 1]->data;
-                            ggml_cuda_op_fused_add(*cuda_ctx, &fused_add_node, n_fuse);
                             i += n_fuse - 1;
 
                             continue;
@@ -4343,6 +4429,8 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .free                    = */ ggml_backend_cuda_free,
     /* .set_tensor_async        = */ ggml_backend_cuda_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_cuda_get_tensor_async,
+    /* .get_tensor_2d_async     = */ ggml_backend_cuda_set_tensor_2d_async,
+    /* .set_tensor_2d_async     = */ ggml_backend_cuda_get_tensor_2d_async,
     /* .cpy_tensor_async        = */ ggml_backend_cuda_cpy_tensor_async,
     /* .synchronize             = */ ggml_backend_cuda_synchronize,
     /* .graph_plan_create       = */ NULL,
@@ -5130,6 +5218,9 @@ static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t
 
 static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
     GGML_UNUSED(reg);
+    if (strcmp(name, "ggml_backend_allreduce_tensor") == 0) {
+        return (void *)ggml_backend_cuda_allreduce_tensor;
+    }
     if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
         return (void *)ggml_backend_cuda_split_buffer_type;
     }

ggml/src/ggml-cuda/mma.cuh

@@ -1025,7 +1025,8 @@ namespace ggml_cuda_mma {
         const floatx2_t& a_frag = reinterpret_cast<const floatx2_t&>(A.x[0]);
         const floatx2_t& b_frag = reinterpret_cast<const floatx2_t&>(B.x[0]);
         acc_frag = __builtin_amdgcn_mfma_f32_16x16x8_xf32(a_frag, b_frag, acc_frag, 0, 0, 0);
-#elif defined(CDNA2) || defined(CDNA1)
+#elif defined(CDNA4) || defined(CDNA2) || defined(CDNA1)
+        // CDNA4 (gfx950) does not support xf32 MFMA, use f32 path like CDNA2/CDNA1
 #pragma unroll
         for (int i = 0; i < 2; ++i) {
             acc_frag = __builtin_amdgcn_mfma_f32_16x16x4f32(A.x[i], B.x[i], acc_frag, 0, 0, 0);
@@ -1187,7 +1188,7 @@ namespace ggml_cuda_mma {
 #elif defined(AMD_MFMA_AVAILABLE)
         using floatx4_t = __attribute__((ext_vector_type(4))) float;
         floatx4_t& acc_frag = reinterpret_cast<floatx4_t&>(D.x[0]);
-#if defined(CDNA3) || defined(CDNA2)
+#if defined(CDNA4) || defined(CDNA3) || defined(CDNA2)
         using bf16x4_t = __attribute__((ext_vector_type(4))) __bf16;
         const bf16x4_t& a_frag = reinterpret_cast<const bf16x4_t&>(A.x[0]);
         const bf16x4_t& b_frag = reinterpret_cast<const bf16x4_t&>(B.x[0]);
@@ -1216,12 +1217,12 @@ namespace ggml_cuda_mma {
 #if defined(AMD_MFMA_AVAILABLE)
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * acc = (int32x4_t *) D.x;
-#if defined(CDNA3)
+#if defined(CDNA4) || defined(CDNA3)
         acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(((int64_t *) A.x)[0],
                                                        ((int64_t *) B.x)[0],
                                                        acc[0],
                                                        0, 0, 0);
-#elif defined(CDNA2) || defined(CDNA)
+#elif defined(CDNA2) || defined(CDNA1)
         acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0],
                                                       B.x[0],
                                                       acc[0],
@@ -1230,7 +1231,7 @@ namespace ggml_cuda_mma {
                                                       B.x[1],
                                                       acc[0],
                                                       0, 0, 0);
-#endif // defined(CDNA3)
+#endif // defined(CDNA4) || defined(CDNA3)
 
 #elif defined(AMD_WMMA_AVAILABLE)
 
@@ -1295,12 +1296,12 @@ namespace ggml_cuda_mma {
 #if defined(AMD_MFMA_AVAILABLE)
         using int32x16_t = __attribute__((__vector_size__(16 * sizeof(int)))) int;
         int32x16_t * acc = (int32x16_t *) D.x;
-#if defined(CDNA3)
+#if defined(CDNA4) || defined(CDNA3)
         acc[0] = __builtin_amdgcn_mfma_i32_32x32x16_i8(((int64_t *) A.x)[0],
                                                        ((int64_t *) B.x)[0],
                                                        acc[0],
                                                        0, 0, 0);
-#elif defined(CDNA2) || defined(CDNA)
+#elif defined(CDNA2) || defined(CDNA1)
         acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[0],
                                                      B.x[0],
                                                      acc[0],
@@ -1309,7 +1310,7 @@ namespace ggml_cuda_mma {
                                                      B.x[1],
                                                      acc[0],
                                                      0, 0, 0);
-#endif // defined(CDNA3)
+#endif // defined(CDNA4) || defined(CDNA3)
 
 #else
         GGML_UNUSED_VARS(D, A, B);

ggml/src/ggml-cuda/mmq.cuh

@@ -3645,7 +3645,7 @@ static __global__ void mul_mat_q(
              tile_x_max_i, tile_y_max_j, 0, ncols_x/qk);
         return;
     }
-#endif // (defined(GGML_USE_HIP) && !defined(CDNA3)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
+#endif // (defined(GGML_USE_HIP) && !defined(CDNA4) && !defined(CDNA3)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
 
     constexpr int ITER_K = get_iter_k(type);
 

ggml/src/ggml-cuda/top-k.cu

@@ -25,14 +25,14 @@ static void top_k_cub(ggml_cuda_pool & pool,
     auto indexes_in = cuda::make_counting_iterator(0);
 
     size_t temp_storage_bytes = 0;
-    DeviceTopK::MaxPairs(nullptr, temp_storage_bytes, src, cuda::discard_iterator(), indexes_in, dst, ncols, k,
-                         env);
+    CUDA_CHECK(DeviceTopK::MaxPairs(nullptr, temp_storage_bytes, src, cuda::discard_iterator(), indexes_in, dst, ncols, k,
+                         env));
 
     ggml_cuda_pool_alloc<uint8_t> temp_storage_alloc(pool, temp_storage_bytes);
     void *                        d_temp_storage = temp_storage_alloc.get();
 
-    DeviceTopK::MaxPairs(d_temp_storage, temp_storage_bytes, src, cuda::discard_iterator(), indexes_in, dst,
-                         ncols, k, env);
+    CUDA_CHECK(DeviceTopK::MaxPairs(d_temp_storage, temp_storage_bytes, src, cuda::discard_iterator(), indexes_in, dst,
+                         ncols, k, env));
 }
 
 #elif defined(GGML_CUDA_USE_CUB)  // CUB_TOP_K_AVAILABLE

ggml/src/ggml-cuda/vendors/cuda.h

@@ -6,6 +6,10 @@
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
 
+#ifdef GGML_USE_NCCL
+#include <nccl.h>
+#endif // GGML_USE_NCCL
+
 #if CUDART_VERSION >= 11080
 #include <cuda_fp8.h>
 #define FP8_AVAILABLE

ggml/src/ggml-cuda/vendors/hip.h

@@ -10,6 +10,11 @@
 #include <rocwmma/rocwmma-version.hpp>
 #endif // defined(GGML_HIP_ROCWMMA_FATTN)
 
+#ifdef GGML_USE_NCCL
+#include <rccl/rccl.h>
+#endif // GGML_USE_NCCL
+
+
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_OP_N HIPBLAS_OP_N
@@ -28,6 +33,7 @@
 #define CU_MEM_LOCATION_TYPE_DEVICE hipMemLocationTypeDevice
 #define CU_MEM_ACCESS_FLAGS_PROT_READWRITE hipMemAccessFlagsProtReadWrite
 #define CU_CHECK(fn) {hipError_t err = fn; if(err != hipSuccess) { GGML_ABORT("HipVMM Failure: %s\n", hipGetErrorString(err)); }}
+#define NCCL_CHECK(fn) {ncclResult_t err = fn; if(err != ncclSuccess) { GGML_ABORT("RCCL Failure RCCL returned: %i\n", err); }}
 #define __shfl_sync(mask, var, laneMask, width) __shfl(var, laneMask, width)
 #define __shfl_up_sync(mask, var, laneMask, width) __shfl_up(var, laneMask, width)
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
@@ -183,6 +189,10 @@
 #define GCN
 #endif // defined(GCN5) || defined(GCN4)
 
+#if defined(__gfx950__)
+#define CDNA4
+#endif // defined(__gfx950__)
+
 #if defined(__gfx942__)
 #define CDNA3
 #endif // defined(__gfx942__)
@@ -195,9 +205,9 @@
 #define CDNA1
 #endif // defined(__gfx908__)
 
-#if defined(CDNA3) || defined(CDNA2) || defined(CDNA1)
+#if defined(CDNA4) || defined(CDNA3) || defined(CDNA2) || defined(CDNA1)
 #define CDNA // For the entire family
-#endif // defined(CDNA3) || defined(CDNA2) || defined(CDNA1)
+#endif // defined(CDNA4) || defined(CDNA3) || defined(CDNA2) || defined(CDNA1)
 
 #if defined(__GFX12__)
 #define RDNA4

ggml/src/ggml-ext.h

@@ -0,0 +1,56 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+// This is a "staging" header for new ggml API
+// It is not publicly available and it should not be used by 3rd party projects
+//
+// When the API matures enough, it will be moved to the official public API
+
+//
+// Meta backend
+//
+
+#define GGML_BACKEND_META_MAX_DEVICES 16
+
+enum ggml_backend_meta_split_axis {
+    // tensor split by tensor dimensions:
+    GGML_BACKEND_SPLIT_AXIS_0   =  0,
+    GGML_BACKEND_SPLIT_AXIS_1   =  1,
+    GGML_BACKEND_SPLIT_AXIS_2   =  2,
+    GGML_BACKEND_SPLIT_AXIS_3   =  3,
+
+    GGML_BACKEND_SPLIT_AXIS_MIRRORED = 10, // all values on all backends
+    GGML_BACKEND_SPLIT_AXIS_PARTIAL  = 11, // each backend has a partial sum
+
+    // for internal bookkeeping only:
+    GGML_BACKEND_SPLIT_AXIS_NONE     = 98,
+    GGML_BACKEND_SPLIT_AXIS_UNKNOWN  = 99,
+};
+GGML_API const char * ggml_backend_meta_split_axis_name(enum ggml_backend_meta_split_axis split_axis);
+
+struct ggml_backend_meta_split_state {
+    enum ggml_backend_meta_split_axis axis;
+
+    // for tensors with axis >= 0 && axis < GGML_MAX_DIMS:
+    //   - each device has a slice of the tensor along the split axis
+    //   - most tensors have n_segments == 1 and a contiguous slice of the tensor data
+    //   - some tensors have an inhomogenenous data layout along the split axis,
+    //     those tensors are divided into segments which are each individually split across devices
+    //   - ne has one entry per segment and device that add up to ggml_tensor::ne for that axis,
+    //     the outer/inner loops are over segments/devices like [seg0_dev0, seg0_dev1, seg1_dev0, seg1_dev1],
+    //   - for example, a transformer may have a fused QKV matrix rather than 3 matrices, those would be 3 separate segments
+    //     that each need to be split individually across devices so that each device gets a slice of Q, K, and V
+    int64_t  ne[16*GGML_BACKEND_META_MAX_DEVICES];
+    uint32_t n_segments;
+};
+
+// function to assign split states for statically allocated tensors, compute tensor split states will be assigned to be compatible:
+typedef struct ggml_backend_meta_split_state(*ggml_backend_meta_get_split_state_t)(const struct ggml_tensor * tensor, void * userdata);
+
+// create a new meta device from "simple" devices, meta buffer type/buffer/backend is then derived from this:
+// TODO: this looks a bit strange - a backend API creates a device. I think we should try
+//       express this as a backend registry functionality instead
+GGML_API ggml_backend_dev_t ggml_backend_meta_device(
+    ggml_backend_dev_t * devs, size_t n_devs, ggml_backend_meta_get_split_state_t get_split_state, void * get_split_state_ud);

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -1491,6 +1491,8 @@ static ggml_backend_buffer_i ggml_backend_hexagon_buffer_interface = {
     /* .memset_tensor   = */ NULL,
     /* .set_tensor      = */ ggml_backend_hexagon_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_hexagon_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ ggml_backend_hexagon_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_hexagon_buffer_clear,
     /* .reset           = */ NULL,
@@ -3002,6 +3004,8 @@ static struct ggml_backend_i hexagon_backend_i = {
     /* .free                    = */ ggml_backend_hexagon_free,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
     /* .synchronize             = */ ggml_backend_hexagon_synchronize,
     /* .graph_plan_create       = */ NULL,

ggml/src/ggml-hip/CMakeLists.txt

@@ -47,6 +47,10 @@ find_package(hip     REQUIRED)
 find_package(hipblas REQUIRED)
 find_package(rocblas REQUIRED)
 
+if (GGML_HIP_RCCL)
+    find_package(rccl REQUIRED)
+endif()
+
 if (${hip_VERSION} VERSION_LESS 6.1)
     message(FATAL_ERROR "At least ROCM/HIP V6.1 is required")
 endif()
@@ -118,6 +122,10 @@ if (NOT GGML_HIP_MMQ_MFMA)
     add_compile_definitions(GGML_HIP_NO_MMQ_MFMA)
 endif()
 
+if (GGML_HIP_RCCL)
+    add_compile_definitions(GGML_USE_NCCL) # RCCL has the same interface as NCCL.
+endif()
+
 if (GGML_HIP_EXPORT_METRICS)
     set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -Rpass-analysis=kernel-resource-usage --save-temps")
 endif()
@@ -142,4 +150,8 @@ if (GGML_STATIC)
     message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
 endif()
 
+if (GGML_HIP_RCCL)
+    target_link_libraries(ggml-hip PRIVATE ggml-base roc::rccl)
+endif()
+
 target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas)

ggml/src/ggml-metal/ggml-metal.cpp

@@ -90,6 +90,8 @@ static ggml_backend_buffer_i ggml_backend_metal_buffer_shared_i = {
     /* .memset_tensor   = */ ggml_backend_metal_buffer_shared_memset_tensor,
     /* .set_tensor      = */ ggml_backend_metal_buffer_shared_set_tensor,
     /* .get_tensor      = */ ggml_backend_metal_buffer_shared_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ ggml_backend_metal_buffer_shared_cpy_tensor,
     /* .clear           = */ ggml_backend_metal_buffer_shared_clear,
     /* .reset           = */ NULL,
@@ -158,15 +160,17 @@ static void ggml_backend_metal_buffer_private_clear(ggml_backend_buffer_t buffer
 }
 
 static ggml_backend_buffer_i ggml_backend_metal_buffer_private_i = {
-    /* .free_buffer     = */ ggml_backend_metal_buffer_private_free_buffer,
-    /* .get_base        = */ ggml_backend_metal_buffer_private_get_base,
-    /* .init_tensor     = */ NULL,
-    /* .memset_tensor   = */ ggml_backend_metal_buffer_private_memset_tensor,
-    /* .set_tensor      = */ ggml_backend_metal_buffer_private_set_tensor,
-    /* .get_tensor      = */ ggml_backend_metal_buffer_private_get_tensor,
-    /* .cpy_tensor      = */ ggml_backend_metal_buffer_private_cpy_tensor,
-    /* .clear           = */ ggml_backend_metal_buffer_private_clear,
-    /* .reset           = */ NULL,
+    /* .free_buffer             = */ ggml_backend_metal_buffer_private_free_buffer,
+    /* .get_base                = */ ggml_backend_metal_buffer_private_get_base,
+    /* .init_tensor             = */ NULL,
+    /* .memset_tensor           = */ ggml_backend_metal_buffer_private_memset_tensor,
+    /* .set_tensor              = */ ggml_backend_metal_buffer_private_set_tensor,
+    /* .get_tensor              = */ ggml_backend_metal_buffer_private_get_tensor,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
+    /* .cpy_tensor              = */ ggml_backend_metal_buffer_private_cpy_tensor,
+    /* .clear                   = */ ggml_backend_metal_buffer_private_clear,
+    /* .reset                   = */ NULL,
 };
 
 static bool ggml_backend_buffer_is_metal(ggml_backend_buffer_t buffer) {
@@ -563,6 +567,8 @@ static ggml_backend_i ggml_backend_metal_i = {
     /* .free                    = */ ggml_backend_metal_free,
     /* .set_tensor_async        = */ ggml_backend_metal_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_metal_get_tensor_async,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ ggml_backend_metal_cpy_tensor_async, // only needed for multi-GPU setups
     /* .synchronize             = */ ggml_backend_metal_synchronize,
     /* .graph_plan_create       = */ NULL,

ggml/src/ggml-metal/ggml-metal.metal

@@ -10079,6 +10079,7 @@ template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mul_mm_id kernel_m
 #if defined(GGML_METAL_HAS_BF16)
 template [[host_name("kernel_mul_mm_id_bf16_f32")]]    kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
 #endif
+template [[host_name("kernel_mul_mm_id_q1_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
 template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
 template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
 template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
@@ -10102,6 +10103,7 @@ template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mul_mm_id kernel_m
 
 template [[host_name("kernel_mul_mm_id_f32_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
 template [[host_name("kernel_mul_mm_id_f16_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q1_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
 template [[host_name("kernel_mul_mm_id_q4_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
 template [[host_name("kernel_mul_mm_id_q4_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
 template [[host_name("kernel_mul_mm_id_q5_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -4063,6 +4063,8 @@ static ggml_backend_i ggml_backend_opencl_i = {
     /* .set_tensor_async        = */ NULL,  /* ggml_backend_opencl_set_tensor_async */
     /* .get_tensor_async        = */ NULL,  /* ggml_backend_opencl_get_tensor_async */
     /* .cpy_tensor_async        = */ NULL,  /* ggml_backend_opencl_cpy_tensor_async */
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .synchronize             = */ ggml_backend_opencl_synchronize,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
@@ -5778,6 +5780,8 @@ static ggml_backend_buffer_i ggml_backend_opencl_buffer_interface = {
     /* .memset_tensor   = */ NULL,
     /* .set_tensor      = */ ggml_backend_opencl_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_opencl_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ NULL,
     /* .clear           = */ ggml_backend_opencl_buffer_clear,
     /* .reset           = */ ggml_backend_opencl_buffer_reset,

ggml/src/ggml-openvino/ggml-openvino.cpp

@@ -412,6 +412,8 @@ static const ggml_backend_buffer_i ggml_backend_openvino_buffer_interface = {
     /* .memset_tensor   = */ ggml_backend_openvino_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_openvino_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_openvino_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ ggml_backend_openvino_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_openvino_buffer_clear,
     /* .reset           = */ NULL,
@@ -617,6 +619,8 @@ static const ggml_backend_i ggml_backend_openvino_interface = {
     /* .free                    = */ ggml_backend_openvino_free,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
+    /* .get_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
     /* .synchronize             = */ NULL,
     /* .graph_plan_create       = */ NULL,

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -706,6 +706,8 @@ static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = {
     /* .memset_tensor   = */ NULL,
     /* .set_tensor      = */ ggml_backend_rpc_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_rpc_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ ggml_backend_rpc_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_rpc_buffer_clear,
     /* .reset           = */ NULL,
@@ -894,6 +896,8 @@ static ggml_backend_i ggml_backend_rpc_interface = {
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .synchronize             = */ ggml_backend_rpc_synchronize,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,

ggml/src/ggml-sycl/fattn-tile.cpp

@@ -44,6 +44,10 @@ void ggml_sycl_flash_attn_ext_tile(ggml_backend_sycl_context & ctx, ggml_tensor
             GGML_ASSERT(V->ne[0] == K->ne[0]);
             ggml_sycl_flash_attn_ext_tile_case<256, 256>(ctx, dst);
         } break;
+        case 512: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_sycl_flash_attn_ext_tile_case<512, 512>(ctx, dst);
+        } break;
         case 576: {
             GGML_ASSERT(V->ne[0] == 512);
             ggml_sycl_flash_attn_ext_tile_case<576, 512>(ctx, dst);

ggml/src/ggml-sycl/fattn-tile.hpp

@@ -67,6 +67,12 @@ static constexpr uint32_t ggml_sycl_fattn_tile_get_config_fp16(const int DKQ, co
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  64,  64)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  64,  64)
 
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  2,  64, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512, 32, 256, 2,  64,  64)
+
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
@@ -124,6 +130,12 @@ static constexpr uint32_t ggml_sycl_fattn_tile_get_config_fp32(const int DKQ, co
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32,  64)
 
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  2, 128, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512, 32, 256, 2,  64,  64)
+
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  32,  64)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  32,  64)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  32,  64)
@@ -131,134 +143,6 @@ static constexpr uint32_t ggml_sycl_fattn_tile_get_config_fp32(const int DKQ, co
     return 0;
 }
 
-static constexpr uint32_t ggml_sycl_fattn_tile_get_config_amd(const int DKQ, const int DV, const int ncols) {
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  2,  64, 3,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  4, 128, 3,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  8, 128, 2,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 256, 2, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 256, 2,  64,  64)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  2,  64, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  4, 128, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  8, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 16, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 32, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 64, 256, 2,  32,  48)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  2,  64, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  4, 128, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  8, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 16, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 32, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 64, 256, 2,  32,  56)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  2, 256, 2, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  4, 128, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  8, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 16, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 2,  64,  32)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  2, 256, 2, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  4, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  8, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32, 128)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128,  64)
-
-    return 0;
-}
-
-static constexpr uint32_t ggml_sycl_fattn_tile_get_config_amd_rdna(const int DKQ, const int DV, const int ncols) {
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  2,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  4,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  8, 128, 5, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 128, 5, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 128, 4,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 128, 5,  64,  64)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  2,  64, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  4, 128, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  8, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 16, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 32, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 64, 256, 2,  32,  48)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  2,  64, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  4, 128, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  8, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 16, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 32, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 64, 256, 2,  32,  56)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  2,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  4, 128, 8,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  8, 128, 8,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 16, 256, 3, 128, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 3, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 3,  64,  64)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  2,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  4, 128, 6,  32, 256)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  8, 128, 6,  32, 256)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5,  32, 256)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3,  64, 128)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128,  64)
-
-    return 0;
-}
-
 static constexpr uint32_t ggml_sycl_fattn_tile_get_config(const int DKQ, const int DV, const int ncols, const int cc) {
     if(fast_fp16_available(cc))
         return ggml_sycl_fattn_tile_get_config_fp16(DKQ, DV, ncols);
@@ -1293,6 +1177,16 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_sycl_context & ctx, ggm
             launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
             return;
         }
+        // ncols2=2 and ncols2=1 fallbacks only for cases where ncols=2 config exists (DKQ == DV).
+        // For DKQ == 576, DV == 512 only GQA-optimized variants are implemented.
+        if constexpr (DKQ == DV) {
+            if (use_gqa_opt && gqa_ratio % 2 == 0) {
+                launch_fattn_tile_switch_ncols1<DKQ, DV, 2, use_logit_softcap>(ctx, dst);
+                return;
+            }
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 1, use_logit_softcap>(ctx, dst);
+            return;
+        }
     }
 
     if constexpr (DV <= 256) {
@@ -1347,5 +1241,6 @@ extern DECL_FATTN_TILE_CASE( 96,  96);
 extern DECL_FATTN_TILE_CASE(112, 112);
 extern DECL_FATTN_TILE_CASE(128, 128);
 extern DECL_FATTN_TILE_CASE(256, 256);
+extern DECL_FATTN_TILE_CASE(512, 512);
 extern DECL_FATTN_TILE_CASE(576, 512);
 

ggml/src/ggml-sycl/fattn-vec.hpp

@@ -664,4 +664,11 @@ EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_0)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_1)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q8_0)
 
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_F16)
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_Q4_0)
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_Q4_1)
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_Q5_0)
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_Q5_1)
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_Q8_0)
+
 #endif // GGML_SYCL_FATTN_VEC_HPP

ggml/src/ggml-sycl/fattn.cpp

@@ -34,6 +34,7 @@
     FATTN_VEC_CASE( 64, type_K, type_V)       \
     FATTN_VEC_CASE(128, type_K, type_V)       \
     FATTN_VEC_CASE(256, type_K, type_V)       \
+    FATTN_VEC_CASE(512, type_K, type_V)       \
 
 static void ggml_sycl_flash_attn_ext_vec(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_tensor * Q = dst->src[0];
@@ -141,6 +142,7 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
         case 128:
         case 112:
         case 256:
+        case 512:
             if (V->ne[0] != K->ne[0]) {
                 return BEST_FATTN_KERNEL_NONE;
             }
@@ -185,7 +187,7 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
     }
 
     // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
-    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    const bool can_use_vector_kernel = Q->ne[0] <= 512 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
     // Todo: Use the XMX kernel if possible:
 

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -411,11 +411,22 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
         assert(tensor->view_src->buffer->buft == buffer->buft);
         return GGML_STATUS_SUCCESS;
     }
-    if ((tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q8_0 || tensor->type == GGML_TYPE_Q4_K || tensor->type == GGML_TYPE_Q6_K) &&
-        !g_ggml_sycl_disable_optimize) {
-        ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
-        tensor->extra                 = extra;
-        ctx->tensor_extras.push_back(extra);  //used to release it when destroy ctx.
+
+    if (!g_ggml_sycl_disable_optimize) {
+        // set reorder extra buffer based on supported type
+        switch (tensor->type) {
+            case GGML_TYPE_Q4_0:
+            case GGML_TYPE_Q8_0:
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q6_K:{
+                ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
+                tensor->extra                 = extra;
+                ctx->tensor_extras.push_back(extra);
+                break;
+            }
+            default:
+                break;
+        }
     }
 
     if (ggml_is_quantized(tensor->type)) {
@@ -627,6 +638,8 @@ static const ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = {
     /* .memset_tensor   = */ ggml_backend_sycl_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_sycl_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_sycl_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ ggml_backend_sycl_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_sycl_buffer_clear,
     /* .reset           = */ ggml_backend_sycl_buffer_reset,
@@ -1073,6 +1086,8 @@ static struct ggml_backend_buffer_i ggml_backend_sycl_split_buffer_interface = {
     /* .memset_tensor   = */ NULL,
     /* .set_tensor      = */ ggml_backend_sycl_split_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_sycl_split_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ NULL,
     /* .clear           = */ ggml_backend_sycl_split_buffer_clear,
     /* .reset           = */ NULL,
@@ -4542,6 +4557,8 @@ static ggml_backend_i ggml_backend_sycl_interface = {
     /* .free                    = */ ggml_backend_sycl_free,
     /* .set_tensor_async        = */ ggml_backend_sycl_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_sycl_get_tensor_async,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ NULL, // ggml_backend_sycl_cpy_tensor_async,
                                            // // TODO: update for the new
                                            // interface

ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp

@@ -0,0 +1,6 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.hpp"
+
+DECL_FATTN_TILE_CASE(512, 512);
+

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_F16);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q4_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_Q4_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q4_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_Q4_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q5_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_Q5_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q5_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_Q5_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q8_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_Q8_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_F16);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_F16);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_F16);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_F16);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_F16);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_F16);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_F16);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_F16);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_F16);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_F16);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);

ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp

@@ -5,3 +5,4 @@
 DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
 DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
+DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);

ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp

@@ -101,6 +101,8 @@ const ggml_backend_buffer_i ggml_backend_remoting_buffer_interface = {
     /* .memset_tensor   = */ NULL,
     /* .set_tensor      = */ ggml_backend_remoting_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_remoting_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ ggml_backend_remoting_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_remoting_buffer_clear,
     /* .reset           = */ NULL,
@@ -113,6 +115,8 @@ const ggml_backend_buffer_i ggml_backend_remoting_buffer_from_ptr_interface = {
     /* .memset_tensor   = */ NULL,
     /* .set_tensor      = */ ggml_backend_remoting_buffer_set_tensor_from_ptr,
     /* .get_tensor      = */ ggml_backend_remoting_buffer_get_tensor_from_ptr,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ ggml_backend_remoting_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_remoting_buffer_clear,
     /* .reset           = */ NULL,

ggml/src/ggml-virtgpu/ggml-backend.cpp

@@ -34,6 +34,8 @@ static ggml_backend_i ggml_backend_remoting_interface = {
     /* .free                    = */ ggml_backend_remoting_free,
     /* .set_tensor_async        = */ NULL,  // ggml_backend_remoting_set_tensor_async,
     /* .get_tensor_async        = */ NULL,  // ggml_backend_remoting_get_tensor_async,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ NULL,  // ggml_backend_remoting_cpy_tensor_async,
     /* .synchronize             = */ NULL,  // ggml_backend_remoting_synchronize,
     /* .graph_plan_create       = */ NULL,

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -13638,6 +13638,8 @@ static ggml_backend_buffer_i ggml_backend_vk_buffer_interface = {
     /* .memset_tensor   = */ ggml_backend_vk_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_vk_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_vk_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ ggml_backend_vk_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_vk_buffer_clear,
     /* .reset           = */ NULL,
@@ -15096,6 +15098,8 @@ static ggml_backend_i ggml_backend_vk_interface = {
     /* .free                    = */ ggml_backend_vk_free,
     /* .set_tensor_async        = */ ggml_backend_vk_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_vk_get_tensor_async,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ ggml_backend_vk_cpy_tensor_async,
     /* .synchronize             = */ ggml_backend_vk_synchronize,
     /* .graph_plan_create       = */ NULL,

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -3013,6 +3013,8 @@ static ggml_backend_i ggml_backend_webgpu_i = {
     /* .free                    = */ ggml_backend_webgpu_free,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
     /* .synchronize             = */ NULL,
     /* .graph_plan_create       = */ NULL,
@@ -3170,6 +3172,8 @@ static ggml_backend_buffer_i ggml_backend_webgpu_buffer_interface = {
     /* .memset_tensor   = */ ggml_backend_webgpu_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_webgpu_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_webgpu_buffer_get_tensor,
+    /* .set_tensor_2d   = */ NULL,
+    /* .get_tensor_2d   = */ NULL,
     /* .cpy_tensor      = */ NULL,  // TODO: optional, implement this
     /* .clear           = */ ggml_backend_webgpu_buffer_clear,
     /* .reset           = */ NULL,  // TODO: optional, think it coordinates with

ggml/src/ggml-zdnn/ggml-zdnn.cpp

@@ -313,6 +313,8 @@ static ggml_backend_buffer_i ggml_backend_zdnn_buffer_i = {
     /* .memset_tensor = */ ggml_backend_zdnn_buffer_memset_tensor,
     /* .set_tensor    = */ ggml_backend_zdnn_buffer_set_tensor,
     /* .get_tensor    = */ ggml_backend_zdnn_buffer_get_tensor,
+    /* .set_tensor_2d = */ NULL,
+    /* .get_tensor_2d = */ NULL,
     /* .cpy_tensor    = */ NULL,
     /* .clear         = */ ggml_backend_zdnn_buffer_clear,
     /* .reset         = */ NULL,
@@ -417,20 +419,22 @@ static enum ggml_status ggml_backend_zdnn_graph_compute(ggml_backend_t backend,
 }
 
 static ggml_backend_i ggml_backend_zdnn_i = {
-    /* .get_name           = */ ggml_backend_zdnn_name,
-    /* .free               = */ ggml_backend_zdnn_free,
-    /* .set_tensor_async   = */ NULL,
-    /* .get_tensor_async   = */ NULL,
-    /* .cpy_tensor_async   = */ NULL,
-    /* .synchronize        = */ NULL,
-    /* .graph_plan_create  = */ NULL,
-    /* .graph_plan_free    = */ NULL,
-    /* .graph_plan_update  = */ NULL,
-    /* .graph_plan_compute = */ NULL,
-    /* .graph_compute      = */ ggml_backend_zdnn_graph_compute,
-    /* .event_record       = */ NULL,
-    /* .event_wait         = */ NULL,
-    /* .graph_optimize     = */ NULL,
+    /* .get_name               = */ ggml_backend_zdnn_name,
+    /* .free                   = */ ggml_backend_zdnn_free,
+    /* .set_tensor_async       = */ NULL,
+    /* .get_tensor_async       = */ NULL,
+    /* .get_tensor_2d_async    = */ NULL,
+    /* .set_tensor_2d_async    = */ NULL,
+    /* .cpy_tensor_async       = */ NULL,
+    /* .synchronize            = */ NULL,
+    /* .graph_plan_create      = */ NULL,
+    /* .graph_plan_free        = */ NULL,
+    /* .graph_plan_update      = */ NULL,
+    /* .graph_plan_compute     = */ NULL,
+    /* .graph_compute          = */ ggml_backend_zdnn_graph_compute,
+    /* .event_record           = */ NULL,
+    /* .event_wait             = */ NULL,
+    /* .graph_optimize         = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_zdnn_guid(void) {

ggml/src/ggml-zendnn/ggml-zendnn.cpp

@@ -407,6 +407,8 @@ static struct ggml_backend_i ggml_backend_zendnn_i = {
     /* .free                    = */ ggml_backend_zendnn_free,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
+    /* .get_tensor_2d_async     = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
     /* .synchronize             = */ NULL,
     /* .graph_plan_create       = */ NULL,

gguf-py/gguf/constants.py

@@ -4122,6 +4122,7 @@ class VisionProjectorType:
     LIGHTONOCR = "lightonocr"
     COGVLM = "cogvlm"
     JANUS_PRO = "janus_pro"
+    DOTSOCR = "dots_ocr"
     DEEPSEEKOCR = "deepseekocr"
     LFM2A = "lfm2a" # audio
     MUSIC_FLAMINGO = "musicflamingo" # audio

gguf-py/gguf/tensor_mapping.py

@@ -1359,6 +1359,7 @@ class TensorNameMap:
             "visual.merger.mlp.{bid}", # qwen2vl
             "mlp_AR.linear_{bid}", # PaddleOCR-VL
             "merger.mlp.{bid}",
+            "vision_tower.merger.mlp.{bid}", # dots.ocr
             "vit.perceive.proj.{bid}", # HunyuanOCR (proj.0 = conv1, proj.2 = conv2)
         ),
 
@@ -1406,11 +1407,13 @@ class TensorNameMap:
             "siglip2.vision_model.embeddings.patch_embedding",
             "vision_model.radio_model.model.patch_generator.embedder", # Nemotron Nano v2 VL
             "model.vision_tower.patch_embedder.input_proj", # gemma4
+            "vision_tower.patch_embed.patchifier.proj", # dots.ocr
             "vision_model.conv1", # Step3-VL
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_NORM: (
             "visual.post_conv_layernorm", # glm4v
+            "vision_tower.patch_embed.patchifier.norm", # dots.ocr
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_POS: (
@@ -1441,6 +1444,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_ENC_ATTN_QKV: (
             "visual.blocks.{bid}.attn.qkv", # qwen3vl
+            "vision_tower.blocks.{bid}.attn.qkv", # dots.ocr
             "model.vision.transformer.layers.{bid}.attention.query_key_value", # cogvlm
             "model.vision_model.transformer.layers.{bid}.self_attn.qkv_proj", # Deepseek-OCR CLIP
             "vision_tower.encoder.blocks.{bid}.wqkv", # Kimi-K2.5
@@ -1526,6 +1530,7 @@ class TensorNameMap:
             "model.vision_model.transformer.layers.{bid}.layer_norm1", # Deepseek-OCR CLIP
             "siglip2.vision_model.encoder.layers.{bid}.layer_norm1",
             "vision_model.radio_model.model.blocks.{bid}.norm1", # Nemotron Nano v2 VL
+            "vision_tower.blocks.{bid}.norm1", # dots.ocr
             "vision_model.transformer.resblocks.{bid}.ln_1", # Step3-VL
         ),
 
@@ -1547,6 +1552,7 @@ class TensorNameMap:
             "siglip2.vision_model.encoder.layers.{bid}.self_attn.out_proj", # youtuvl
             "vision_model.radio_model.model.blocks.{bid}.attn.proj", # Nemotron Nano v2 VL
             "vision_model.model.layers.{bid}.self_attn.o_proj.linear", # gemma4
+            "vision_tower.blocks.{bid}.attn.proj", # dots.ocr
             "vision_model.transformer.resblocks.{bid}.attn.out_proj", # Step3-VL
         ),
 
@@ -1567,6 +1573,7 @@ class TensorNameMap:
             "siglip2.vision_model.encoder.layers.{bid}.layer_norm2",
             "vision_model.radio_model.model.blocks.{bid}.norm2", # Nemotron Nano v2 VL
             "vision_model.model.layers.{bid}.pre_feedforward_layernorm", # gemma4
+            "vision_tower.blocks.{bid}.norm2", # dots.ocr
             "vision_model.transformer.resblocks.{bid}.ln_2", # Step3-VL
         ),
 
@@ -1649,6 +1656,7 @@ class TensorNameMap:
             "vision_encoder.ln_pre", # pixtral
             "vision_model.layernorm_pre", # llama4
             "model.vision_model.pre_layrnorm", # Deepseek-OCR CLIP
+            "vision_tower.patch_embed.patchifier.norm", # dots.ocr
             "vision_model.ln_pre", # Step3-VL
         ),
 
@@ -1664,6 +1672,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_MM_POST_NORM: (
             "visual.merger.post_projection_norm", # glm4v
+            "vision_tower.post_trunk_norm", # dots.ocr
             "vit.perceive.after_rms", # HunyuanOCR
         ),
 
@@ -1680,6 +1689,7 @@ class TensorNameMap:
             "model.vision.linear_proj.norm1", # cogvlm
             "mlp_AR.pre_norm", # PaddleOCR-VL
             "merger.ln_q",
+            "vision_tower.merger.ln_q", # dots.ocr
         ),
 
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM: (

gguf-py/gguf/vocab.py

@@ -543,7 +543,7 @@ class LlamaHfVocab(Vocab):
             cache_dir=base_path,
             local_files_only=True,
         )
-        assert self.tokenizer.is_fast  # assume tokenizer.json is used
+        assert self.tokenizer.is_fast  # assume tokenizer.json is used  # ty: ignore[unresolved-attribute]
 
         # Initialize lists and dictionaries for added tokens
         self.added_tokens_list = []
@@ -552,30 +552,30 @@ class LlamaHfVocab(Vocab):
 
         # Process added tokens
         for tok, tokidx in sorted(
-            self.tokenizer.get_added_vocab().items(), key=lambda x: x[1]
+            self.tokenizer.get_added_vocab().items(), key=lambda x: x[1]  # ty: ignore[unresolved-attribute]
         ):
             # Only consider added tokens that are not in the base vocabulary
-            if tokidx >= self.tokenizer.vocab_size:
+            if tokidx >= self.tokenizer.vocab_size:  # ty: ignore[unresolved-attribute]
                 self.added_tokens_list.append(tok)
                 self.added_tokens_dict[tok] = tokidx
                 self.added_tokens_ids.add(tokidx)
 
         # Store special tokens and their IDs
         self.specials = {
-            tok: self.tokenizer.get_vocab()[tok]
-            for tok in self.tokenizer.all_special_tokens
+            tok: self.tokenizer.get_vocab()[tok]  # ty: ignore[unresolved-attribute]
+            for tok in self.tokenizer.all_special_tokens  # ty: ignore[unresolved-attribute]
         }
-        self.special_ids = set(self.tokenizer.all_special_ids)
+        self.special_ids = set(self.tokenizer.all_special_ids)  # ty: ignore[unresolved-attribute]
 
         # Set vocabulary sizes
-        self.vocab_size_base = self.tokenizer.vocab_size
+        self.vocab_size_base = self.tokenizer.vocab_size  # ty: ignore[unresolved-attribute]
         self.vocab_size      = self.vocab_size_base + len(self.added_tokens_list)
 
         self.fname_tokenizer = fname_tokenizer
 
     def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         reverse_vocab = {
-            id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items()
+            id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items()  # ty: ignore[unresolved-attribute]
         }
 
         for token_id in range(self.vocab_size_base):
@@ -616,7 +616,7 @@ class LlamaHfVocab(Vocab):
             yield text.encode("utf-8"), score, toktype
 
     def has_newline_token(self):
-        return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab
+        return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab  # ty: ignore[unresolved-attribute]
 
     def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         yield from self.hf_tokens()

include/llama.h

@@ -192,9 +192,10 @@ extern "C" {
     LLAMA_API const char * llama_flash_attn_type_name(enum llama_flash_attn_type flash_attn_type);
 
     enum llama_split_mode {
-        LLAMA_SPLIT_MODE_NONE  = 0, // single GPU
-        LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
-        LLAMA_SPLIT_MODE_ROW   = 2, // split layers and KV across GPUs, use tensor parallelism if supported
+        LLAMA_SPLIT_MODE_NONE   = 0, // single GPU
+        LLAMA_SPLIT_MODE_LAYER  = 1, // split layers and KV across GPUs
+        LLAMA_SPLIT_MODE_ROW    = 2, // split layers and KV across GPUs, use tensor parallelism if supported
+        LLAMA_SPLIT_MODE_TENSOR = 3,
     };
 
     // TODO: simplify (https://github.com/ggml-org/llama.cpp/pull/9294#pullrequestreview-2286561979)

models/ggml-vocab-gemma-4.gguf.inp

@@ -0,0 +1,111 @@
+ied 4 ½ months
+__ggml_vocab_test__
+Äpfel
+__ggml_vocab_test__
+
+__ggml_vocab_test__
+ 
+__ggml_vocab_test__
+  
+__ggml_vocab_test__
+   
+__ggml_vocab_test__
+	
+__ggml_vocab_test__
+
+
+__ggml_vocab_test__
+
+
+
+__ggml_vocab_test__
+
+
+
+
+__ggml_vocab_test__
+	
+
+__ggml_vocab_test__
+Hello world
+__ggml_vocab_test__
+ Hello world
+__ggml_vocab_test__
+Hello World
+__ggml_vocab_test__
+ Hello World
+__ggml_vocab_test__
+ Hello World!
+__ggml_vocab_test__
+Hello, world!
+__ggml_vocab_test__
+ Hello, world!
+__ggml_vocab_test__
+ this is 🦙.cpp
+__ggml_vocab_test__
+w048 7tuijk dsdfhu
+__ggml_vocab_test__
+нещо на Български
+__ggml_vocab_test__
+កាន់តែពិសេសអាចខលចេញ
+__ggml_vocab_test__
+🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)
+__ggml_vocab_test__
+Hello
+__ggml_vocab_test__
+ Hello
+__ggml_vocab_test__
+  Hello
+__ggml_vocab_test__
+   Hello
+__ggml_vocab_test__
+    Hello
+__ggml_vocab_test__
+    Hello
+    Hello
+__ggml_vocab_test__
+ (
+__ggml_vocab_test__
+
+ =
+__ggml_vocab_test__
+' era
+__ggml_vocab_test__
+Hello, y'all! How are you 😁 ?我想在apple工作1314151天～
+__ggml_vocab_test__
+!!!!!!
+__ggml_vocab_test__
+3
+__ggml_vocab_test__
+33
+__ggml_vocab_test__
+333
+__ggml_vocab_test__
+3333
+__ggml_vocab_test__
+33333
+__ggml_vocab_test__
+333333
+__ggml_vocab_test__
+3333333
+__ggml_vocab_test__
+33333333
+__ggml_vocab_test__
+333333333
+__ggml_vocab_test__
+Cửa Việt
+__ggml_vocab_test__
+ discards
+__ggml_vocab_test__
+
+ 
+
+ 
+
+
+ 	 		 	
+  
+   
+    
+     
+🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL

models/ggml-vocab-gemma-4.gguf.out

@@ -0,0 +1,46 @@
+1178 236743 236812 47041 3794
+239122 22744 535
+
+236743
+138
+139
+255968
+107
+108
+109
+255968 107
+9259 1902
+26352 1902
+9259 4109
+26352 4109
+26352 4109 236888
+9259 236764 1902 236888
+26352 236764 1902 236888
+672 563 236743 478 397 404 391 236761 12362
+236765 236771 236812 236828 236743 236832 11372 12065 31806 3405 9360
+1337 12515 1333 4632 165543 3830
+234889 63031 219876 66212 239077 237907 144494
+242015 568 7382 236768 236743 247717 237243 248989 238178 568 43819 111730 150567 236768 113452 568 8960 64334 600 815 1061 1852 8369 236768
+9259
+26352
+138 9259
+139 9259
+140 9259
+140 9259 107 140 9259
+568
+107 578
+236789 6933
+9259 236764 570 236789 712 236888 2088 659 611 170124 2360 62133 237075 17641 11700 236770 236800 236770 236812 236770 236810 236770 237471 238352
+123947
+236800
+236800 236800
+236800 236800 236800
+236800 236800 236800 236800
+236800 236800 236800 236800 236800
+236800 236800 236800 236800 236800 236800
+236800 236800 236800 236800 236800 236800 236800
+236800 236800 236800 236800 236800 236800 236800 236800
+236800 236800 236800 236800 236800 236800 236800 236800 236800
+236780 29719 33154
+2243 2206
+107 236743 108 236743 109 236743 255968 236743 255969 236743 255968 107 138 107 139 107 140 107 141 107 242015 568 7382 236768 236743 247717 237243 248989 238178 568 43819 111730 150567 236768 113452 236743 478 397 404 391 478 397 404 391 236743 236800 236743 236800 236800 236743 236800 236800 236800 236743 236800 236800 236800 236800 236743 236800 236800 236800 236800 236800 236743 236800 236800 236800 236800 236800 236800 236743 236800 236800 236800 236800 236800 236800 236800 236743 236800 236800 236800 236800 236800 236800 236800 236800 236743 236800 236761 236800 236743 236800 856 236800 236743 236800 1390 236800 90986 92814 63031 219876 66212 241702 2360 62133 237075 17641 11700 236770 236800 236770 236812 236770 236810 236770 237471 238352 80448 120697 210119 1333 4632 165543 3830 9451 159561 2629 2629 2717 84491 19938 123947 38950 10371 564 236789 560 1010 756 151812 668 236789 236751 993 236764 756 1357 611 2889 236881 756 236792 711 2889 564 236789 859 1386 625 236764 756 236796 611 1133 1070 11115 236881 1191 236789 32541 496 236789 95635

pyproject.toml

@@ -18,7 +18,7 @@ classifiers = [
 python = ">=3.9"
 numpy = "^1.25.0"
 sentencepiece = ">=0.1.98,<0.3.0"
-transformers = ">=4.35.2,<5.0.0"
+transformers = "==5.5.1"
 protobuf = ">=4.21.0,<5.0.0"
 gguf = { path = "./gguf-py" }
 torch = { version = "^2.2.0", source = "pytorch" }