diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 0951469149..61eb9839e0 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -1261,6 +1261,9 @@ class TextModel(ModelBase): if chkhsh == "6c81ce329e0802883b22eabab0d3fa48357337ef1ecb45443828bf1f6254833f": # ref: https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B res = "exaone-moe" + if chkhsh == "d30d75d9059f1aa2c19359de71047b3ae408c70875e8a3ccf8c5fba56c9d8af4": + # ref: https://huggingface.co/Qwen/Qwen3.5-9B-Instruct + res = "qwen35" if res is None: logger.warning("\n") @@ -4287,6 +4290,7 @@ class Qwen3NextModel(Qwen2MoeModel): self.gguf_writer.add_ssm_group_count(self.hparams["linear_num_key_heads"]) self.gguf_writer.add_ssm_time_step_rank(self.hparams["linear_num_value_heads"]) self.gguf_writer.add_ssm_inner_size(self.hparams["linear_value_head_dim"] * self.hparams["linear_num_value_heads"]) + self.gguf_writer.add_full_attention_interval(self.hparams.get("full_attention_interval", 4)) if (rope_dim := self.hparams.get("head_dim")) is None: rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"] self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.25))) @@ -4351,7 +4355,7 @@ class RND1Model(Qwen2MoeModel): self.gguf_writer.add_mask_token_id(mask_token_id) -@ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration") +@ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration", "Qwen3_5ForConditionalGeneration", "Qwen3_5MoeForConditionalGeneration") class Qwen3VLVisionModel(MmprojModel): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) @@ -4397,6 +4401,10 @@ class Qwen3VLVisionModel(MmprojModel): if name.startswith("model.language_model.") or name.startswith("lm_head."): return + # Skip MTP tensors + if name.startswith("mtp."): + return + if name.startswith("model.visual."): name = name.replace("model.visual.", "visual.", 1) @@ -4559,6 +4567,93 @@ class Qwen3VLMoeTextModel(Qwen3MoeModel): yield from super().modify_tensors(data_torch, name, bid) +class _LinearAttentionVReorderBase(Qwen3NextModel): + model_arch = gguf.MODEL_ARCH.QWEN3NEXT # overridden by subclasses + """reorders V heads from grouped to tiled order for ggml broadcast + + see https://github.com/ggml-org/llama.cpp/pull/19468#discussion_r2786394306 + + Linear attention may has num_k_heads < num_v_heads. The HF weights store + V heads grouped by K head: [G0_v0..v{r-1}, G1_v0..v{r-1}, ...]. + ggml binary ops use tiled broadcast: [K0, K1, ..., K0, K1, ...]. + We reorder V heads to tiled order so ggml_repeat can replace the expensive + interleaved repeat: [G0_v0, G1_v0, ..., G0_v1, G1_v1, ...]. + """ + + @staticmethod + def _reorder_v_heads(tensor: Tensor, dim: int, num_k_heads: int, num_v_per_k: int, head_dim: int) -> Tensor: + """Reorder V heads from grouped (by K head) to tiled order along the given dimension.""" + shape = list(tensor.shape) + if dim < 0: + dim += len(shape) + new_shape = shape[:dim] + [num_k_heads, num_v_per_k, head_dim] + shape[dim + 1:] + tensor = tensor.reshape(*new_shape) + perm = list(range(len(new_shape))) + perm[dim], perm[dim + 1] = perm[dim + 1], perm[dim] + return tensor.permute(*perm).contiguous().reshape(*shape) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + num_k_heads = self.hparams.get("linear_num_key_heads", 0) + num_v_heads = self.hparams.get("linear_num_value_heads", 0) + + if num_k_heads > 0 and num_v_heads > 0 and num_k_heads != num_v_heads and "linear_attn." in name: + head_k_dim = self.hparams["linear_key_head_dim"] + head_v_dim = self.hparams["linear_value_head_dim"] + num_v_per_k = num_v_heads // num_k_heads + + if ".in_proj_qkv." in name: + # QKV weight: reorder only the V rows + q_dim = head_k_dim * num_k_heads + k_dim = head_k_dim * num_k_heads + q = data_torch[:q_dim] + k = data_torch[q_dim:q_dim + k_dim] + v = data_torch[q_dim + k_dim:] + v = self._reorder_v_heads(v, 0, num_k_heads, num_v_per_k, head_v_dim) + data_torch = torch.cat([q, k, v], dim=0) + + elif ".in_proj_z." in name: + # Z gate weight: reorder rows (num_v_heads * head_v_dim) + data_torch = self._reorder_v_heads(data_torch, 0, num_k_heads, num_v_per_k, head_v_dim) + + elif ".in_proj_b." in name or ".in_proj_a." in name: + # Beta/Alpha weight: reorder rows (num_v_heads, head_dim=1) + data_torch = self._reorder_v_heads(data_torch, 0, num_k_heads, num_v_per_k, 1) + + elif ".A_log" in name or ".dt_bias" in name or ".dt_proj" in name: + # A_log / dt_bias: 1D parameters with num_v_heads elements + if data_torch.ndim == 1: + data_torch = self._reorder_v_heads( + data_torch.unsqueeze(-1), 0, num_k_heads, num_v_per_k, 1 + ).squeeze(-1) + else: + data_torch = self._reorder_v_heads(data_torch, -1, num_k_heads, num_v_per_k, 1) + + elif ".conv1d" in name: + # Conv1d kernel: reorder only the V channel portion + data = data_torch.squeeze() + qk_channels = head_k_dim * num_k_heads * 2 + qk_part = data[:qk_channels] + v_part = data[qk_channels:] + v_part = self._reorder_v_heads(v_part, 0, num_k_heads, num_v_per_k, head_v_dim) + data_torch = torch.cat([qk_part, v_part], dim=0) + + elif ".out_proj." in name: + # Out projection weight: reorder columns (input dimension) + data_torch = self._reorder_v_heads(data_torch, 1, num_k_heads, num_v_per_k, head_v_dim) + + yield from super().modify_tensors(data_torch, name, bid) + + +@ModelBase.register("Qwen3_5ForConditionalGeneration") +class Qwen3_5TextModel(_LinearAttentionVReorderBase): + model_arch = gguf.MODEL_ARCH.QWEN35 + + +@ModelBase.register("Qwen3_5MoeForConditionalGeneration") +class Qwen3_5MoeTextModel(_LinearAttentionVReorderBase): + model_arch = gguf.MODEL_ARCH.QWEN35MOE + + @ModelBase.register("GPT2LMHeadModel") class GPT2Model(TextModel): model_arch = gguf.MODEL_ARCH.GPT2 diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index 2811f7f884..a683451508 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -148,6 +148,7 @@ models = [ {"name": "youtu", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Youtu-LLM-2B", }, {"name": "solar-open", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/upstage/Solar-Open-100B", }, {"name": "exaone-moe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B", }, + {"name": "qwen35", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3.5-9B-Instruct", } ] # some models are known to be broken upstream, so we will skip them as exceptions diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 3af4fffe95..ec36b6292a 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -142,6 +142,7 @@ class Keys: EMBEDDING_SCALE = "{arch}.embedding_scale" TOKEN_SHIFT_COUNT = "{arch}.token_shift_count" INTERLEAVE_MOE_LAYER_STEP = "{arch}.interleave_moe_layer_step" + FULL_ATTENTION_INTERVAL = "{arch}.full_attention_interval" ACTIVATION_SPARSITY_SCALE = "{arch}.activation_sparsity_scale" ALTUP_ACTIVE_IDX = "{arch}.altup.active_idx" ALTUP_NUM_INPUTS = "{arch}.altup.num_inputs" @@ -384,6 +385,8 @@ class MODEL_ARCH(IntEnum): QWEN3NEXT = auto() QWEN3VL = auto() QWEN3VLMOE = auto() + QWEN35 = auto() + QWEN35MOE = auto() PHI2 = auto() PHI3 = auto() PHIMOE = auto() @@ -557,13 +560,14 @@ class MODEL_TENSOR(IntEnum): SSM_D = auto() SSM_NORM = auto() SSM_OUT = auto() + SSM_ALPHA = auto() # qwen3.5 SSM_BETA_ALPHA = auto() # qwen3next SSM_CONV1D_Q = auto() # Kimi Linear SSM_CONV1D_K = auto() # Kimi Linear SSM_CONV1D_V = auto() # Kimi Linear SSM_F_A = auto() # Kimi Linear SSM_F_B = auto() # Kimi Linear - SSM_BETA = auto() # Kimi Linear + SSM_BETA = auto() # Kimi Linear qwen3.5 SSM_G_A = auto() # Kimi Linear SSM_G_B = auto() # Kimi Linear TIME_MIX_W0 = auto() @@ -814,6 +818,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.QWEN3NEXT: "qwen3next", MODEL_ARCH.QWEN3VL: "qwen3vl", MODEL_ARCH.QWEN3VLMOE: "qwen3vlmoe", + MODEL_ARCH.QWEN35: "qwen35", + MODEL_ARCH.QWEN35MOE: "qwen35moe", MODEL_ARCH.PHI2: "phi2", MODEL_ARCH.PHI3: "phi3", MODEL_ARCH.PHIMOE: "phimoe", @@ -985,13 +991,14 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d", MODEL_TENSOR.SSM_NORM: "blk.{bid}.ssm_norm", MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out", + MODEL_TENSOR.SSM_ALPHA: "blk.{bid}.ssm_alpha", # qwen3.5 MODEL_TENSOR.SSM_BETA_ALPHA: "blk.{bid}.ssm_ba", MODEL_TENSOR.SSM_CONV1D_Q: "blk.{bid}.ssm_conv1d_q", # Kimi Linear MODEL_TENSOR.SSM_CONV1D_K: "blk.{bid}.ssm_conv1d_k", # Kimi Linear MODEL_TENSOR.SSM_CONV1D_V: "blk.{bid}.ssm_conv1d_v", # Kimi Linear MODEL_TENSOR.SSM_F_A: "blk.{bid}.ssm_f_a", # Kimi Linear MODEL_TENSOR.SSM_F_B: "blk.{bid}.ssm_f_b", # Kimi Linear - MODEL_TENSOR.SSM_BETA: "blk.{bid}.ssm_beta", # Kimi Linear + MODEL_TENSOR.SSM_BETA: "blk.{bid}.ssm_beta", # Kimi Linear qwen3.5 MODEL_TENSOR.SSM_G_A: "blk.{bid}.ssm_g_a", # Kimi Linear MODEL_TENSOR.SSM_G_B: "blk.{bid}.ssm_g_b", # Kimi Linear MODEL_TENSOR.TIME_MIX_W0: "blk.{bid}.time_mix_w0", @@ -1818,6 +1825,61 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN_EXP, MODEL_TENSOR.FFN_UP_EXP, ], + MODEL_ARCH.QWEN35: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.ATTN_GATE, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.SSM_A, + MODEL_TENSOR.SSM_CONV1D, + MODEL_TENSOR.SSM_DT, + MODEL_TENSOR.SSM_NORM, + MODEL_TENSOR.SSM_BETA, + MODEL_TENSOR.SSM_ALPHA, + MODEL_TENSOR.SSM_OUT + ], + MODEL_ARCH.QWEN35MOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.ATTN_GATE, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_INP_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.SSM_A, + MODEL_TENSOR.SSM_CONV1D, + MODEL_TENSOR.SSM_DT, + MODEL_TENSOR.SSM_NORM, + MODEL_TENSOR.SSM_BETA, + MODEL_TENSOR.SSM_ALPHA, + MODEL_TENSOR.SSM_OUT + ], MODEL_ARCH.PLAMO: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 62172b24c3..a237537c8d 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -708,6 +708,9 @@ class GGUFWriter: def add_leading_dense_block_count(self, length: int) -> None: self.add_uint32(Keys.LLM.LEADING_DENSE_BLOCK_COUNT.format(arch=self.arch), length) + def add_full_attention_interval(self, interval: int) -> None: + self.add_uint32(Keys.LLM.FULL_ATTENTION_INTERVAL.format(arch=self.arch), interval) + def add_feed_forward_length(self, length: int | Sequence[int]) -> None: if isinstance(length, int): self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 167ade7803..71dd7e887b 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -228,6 +228,7 @@ class TensorNameMap: "transformer_encoder.{bid}.qkv", # neobert "layers.{bid}.attn.Wqkv", # modern-bert "model.layers.{bid}.self_attn.language_expert_query_key_value", # cogvlm + "model.layers.{bid}.linear_attn.in_proj_qkv", # qwen3.5 ), # Attention query @@ -359,6 +360,7 @@ class TensorNameMap: MODEL_TENSOR.ATTN_GATE: ( "model.layers.{bid}.self_attn.gate_proj", # afmoe + "model.layers.{bid}.linear_attn.in_proj_z", # qwen3.5 "model.layers.{bid}.self_attn.g_proj", # step3.5 head-wise attention gate ), @@ -823,6 +825,10 @@ class TensorNameMap: "model.layers.layers.{bid}.mixer.out_proj", # plamo2 ), + MODEL_TENSOR.SSM_ALPHA: ( + "model.layers.{bid}.linear_attn.in_proj_a", # qwen3.5 + ), + MODEL_TENSOR.SSM_BETA_ALPHA: ( "model.layers.{bid}.linear_attn.in_proj_ba", # qwen3next ), @@ -844,7 +850,8 @@ class TensorNameMap: "model.layers.{bid}.self_attn.f_b_proj", ), MODEL_TENSOR.SSM_BETA: ( - "model.layers.{bid}.self_attn.b_proj", + "model.layers.{bid}.linear_attn.in_proj_b", # qwen3.5 + "model.layers.{bid}.self_attn.b_proj", # Kimi Linear ), MODEL_TENSOR.SSM_G_A: ( "model.layers.{bid}.self_attn.g_a_proj", diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index 2115fc4255..fdda05d3ea 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -122,6 +122,8 @@ add_library(llama models/qwen3vl-moe.cpp models/qwen3moe.cpp models/qwen3next.cpp + models/qwen35.cpp + models/qwen35moe.cpp models/refact.cpp models/rnd1.cpp models/rwkv6-base.cpp diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index bd78f1e556..a943d40dc4 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -37,6 +37,8 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_QWEN3NEXT, "qwen3next" }, { LLM_ARCH_QWEN3VL, "qwen3vl" }, { LLM_ARCH_QWEN3VLMOE, "qwen3vlmoe" }, + { LLM_ARCH_QWEN35, "qwen35" }, + { LLM_ARCH_QWEN35MOE, "qwen35moe" }, { LLM_ARCH_PHI2, "phi2" }, { LLM_ARCH_PHI3, "phi3" }, { LLM_ARCH_PHIMOE, "phimoe" }, @@ -195,6 +197,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" }, { LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" }, { LLM_KV_INTERLEAVE_MOE_LAYER_STEP, "%s.interleave_moe_layer_step" }, + { LLM_KV_FULL_ATTENTION_INTERVAL, "%s.full_attention_interval" }, { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, @@ -366,6 +369,7 @@ static const std::map LLM_TENSOR_NAMES = { { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" }, { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" }, { LLM_TENSOR_SSM_BETA_ALPHA, "blk.%d.ssm_ba" }, + { LLM_TENSOR_SSM_ALPHA, "blk.%d.ssm_alpha" }, { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" }, { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" }, { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, @@ -968,7 +972,6 @@ static std::set llm_get_tensor_names(llm_arch arch) { LLM_TENSOR_ATTN_OUT, LLM_TENSOR_ATTN_QKV, LLM_TENSOR_ATTN_GATE, - LLM_TENSOR_FFN_NORM, LLM_TENSOR_FFN_GATE_INP, LLM_TENSOR_FFN_GATE_EXPS, LLM_TENSOR_FFN_DOWN_EXPS, @@ -985,6 +988,63 @@ static std::set llm_get_tensor_names(llm_arch arch) { LLM_TENSOR_SSM_NORM, LLM_TENSOR_SSM_OUT, }; + case LLM_ARCH_QWEN35: + return { + LLM_TENSOR_TOKEN_EMBD, + LLM_TENSOR_OUTPUT_NORM, + LLM_TENSOR_OUTPUT, + LLM_TENSOR_ATTN_NORM, + LLM_TENSOR_ATTN_POST_NORM, + LLM_TENSOR_ATTN_Q, + LLM_TENSOR_ATTN_Q_NORM, + LLM_TENSOR_ATTN_K, + LLM_TENSOR_ATTN_K_NORM, + LLM_TENSOR_ATTN_V, + LLM_TENSOR_ATTN_OUT, + LLM_TENSOR_ATTN_QKV, + LLM_TENSOR_ATTN_GATE, + LLM_TENSOR_FFN_GATE, + LLM_TENSOR_FFN_DOWN, + LLM_TENSOR_FFN_UP, + LLM_TENSOR_SSM_A_NOSCAN, + LLM_TENSOR_SSM_CONV1D, + LLM_TENSOR_SSM_DT, + LLM_TENSOR_SSM_BETA, + LLM_TENSOR_SSM_ALPHA, + LLM_TENSOR_SSM_NORM, + LLM_TENSOR_SSM_OUT, + }; + case LLM_ARCH_QWEN35MOE: + return { + LLM_TENSOR_TOKEN_EMBD, + LLM_TENSOR_OUTPUT_NORM, + LLM_TENSOR_OUTPUT, + LLM_TENSOR_ATTN_NORM, + LLM_TENSOR_ATTN_POST_NORM, + LLM_TENSOR_ATTN_Q, + LLM_TENSOR_ATTN_Q_NORM, + LLM_TENSOR_ATTN_K, + LLM_TENSOR_ATTN_K_NORM, + LLM_TENSOR_ATTN_V, + LLM_TENSOR_ATTN_OUT, + LLM_TENSOR_ATTN_QKV, + LLM_TENSOR_ATTN_GATE, + LLM_TENSOR_FFN_GATE_INP, + LLM_TENSOR_FFN_GATE_EXPS, + LLM_TENSOR_FFN_DOWN_EXPS, + LLM_TENSOR_FFN_UP_EXPS, + LLM_TENSOR_FFN_GATE_INP_SHEXP, + LLM_TENSOR_FFN_GATE_SHEXP, + LLM_TENSOR_FFN_DOWN_SHEXP, + LLM_TENSOR_FFN_UP_SHEXP, + LLM_TENSOR_SSM_A_NOSCAN, + LLM_TENSOR_SSM_CONV1D, + LLM_TENSOR_SSM_DT, + LLM_TENSOR_SSM_BETA, + LLM_TENSOR_SSM_ALPHA, + LLM_TENSOR_SSM_NORM, + LLM_TENSOR_SSM_OUT, + }; case LLM_ARCH_QWEN3VL: case LLM_ARCH_CHAMELEON: case LLM_ARCH_HUNYUAN_DENSE: @@ -2456,6 +2516,7 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_SSM_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_SSM_DT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_SSM_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_SSM_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_SSM_BETA_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, @@ -2675,6 +2736,8 @@ bool llm_arch_is_hybrid(const llm_arch & arch) { case LLM_ARCH_NEMOTRON_H_MOE: case LLM_ARCH_QWEN3NEXT: case LLM_ARCH_KIMI_LINEAR: + case LLM_ARCH_QWEN35: + case LLM_ARCH_QWEN35MOE: return true; default: return false; diff --git a/src/llama-arch.h b/src/llama-arch.h index e8263369b8..4f7b51e70d 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -41,6 +41,8 @@ enum llm_arch { LLM_ARCH_QWEN3NEXT, LLM_ARCH_QWEN3VL, LLM_ARCH_QWEN3VLMOE, + LLM_ARCH_QWEN35, + LLM_ARCH_QWEN35MOE, LLM_ARCH_PHI2, LLM_ARCH_PHI3, LLM_ARCH_PHIMOE, @@ -199,6 +201,7 @@ enum llm_kv { LLM_KV_EMBEDDING_SCALE, LLM_KV_TOKEN_SHIFT_COUNT, LLM_KV_INTERLEAVE_MOE_LAYER_STEP, + LLM_KV_FULL_ATTENTION_INTERVAL, LLM_KV_ATTENTION_HEAD_COUNT, LLM_KV_ATTENTION_HEAD_COUNT_KV, @@ -404,13 +407,14 @@ enum llm_tensor { LLM_TENSOR_SSM_NORM, LLM_TENSOR_SSM_OUT, LLM_TENSOR_SSM_BETA_ALPHA, // qwen3next + LLM_TENSOR_SSM_ALPHA, // qwen3.5 // Kimi Linear KDA (using SSM_ prefix for consistency) LLM_TENSOR_SSM_CONV1D_Q, // kimi: Q conv1d weight LLM_TENSOR_SSM_CONV1D_K, // kimi: K conv1d weight LLM_TENSOR_SSM_CONV1D_V, // kimi: V conv1d weight LLM_TENSOR_SSM_F_A, // kimi: forget gate projection A LLM_TENSOR_SSM_F_B, // kimi: forget gate projection B - LLM_TENSOR_SSM_BETA, // kimi: beta mixing coefficient + LLM_TENSOR_SSM_BETA, // kimi: beta mixing coefficient and qwen3.5 LLM_TENSOR_SSM_G_A, // kimi: output gate projection A LLM_TENSOR_SSM_G_B, // kimi: output gate projection B LLM_TENSOR_TIME_MIX_W0, diff --git a/src/llama-context.cpp b/src/llama-context.cpp index a6df893a31..1a35ce369c 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -2013,7 +2013,7 @@ void llama_context::output_reorder() { // uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const { - if (model.arch == LLM_ARCH_QWEN3NEXT || model.arch == LLM_ARCH_KIMI_LINEAR) { + if (model.arch == LLM_ARCH_QWEN3NEXT || model.arch == LLM_ARCH_KIMI_LINEAR || model.arch == LLM_ARCH_QWEN35 || model.arch == LLM_ARCH_QWEN35MOE) { return std::max(n_tokens * 40, 32u * model.n_tensors()); } uint32_t res = std::max(1024u, 8u*model.n_tensors()); diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 674d06c891..7a06e96c87 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -125,6 +125,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_21B_A3B: return "21B.A3B"; case LLM_TYPE_30B_A3B: return "30B.A3B"; case LLM_TYPE_31B_A3_5B: return "31B.A3.5B"; + case LLM_TYPE_35B_A3B: return "35B.A3B"; case LLM_TYPE_48B_A3B: return "48B.A3B"; case LLM_TYPE_80B_A3B: return "80B.A3B"; case LLM_TYPE_100B_A6B: return "100B.A6B"; @@ -2403,8 +2404,12 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); // Mark recurrent layers (linear attention layers) - for (uint32_t i = 0; i < hparams.n_layer; ++i) { - hparams.recurrent_layer_arr[i] = ((i + 1) % 4 != 0); // TODO: extract the magic 4 from "full_attention_interval" + { + uint32_t full_attn_interval = 4; + ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false); + for (uint32_t i = 0; i < hparams.n_layer; ++i) { + hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0); + } } switch (hparams.n_layer) { @@ -2412,6 +2417,62 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_QWEN35: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true); + + // Load linear attention (gated delta net) parameters + ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); + + // Mark recurrent layers (linear attention layers) + { + uint32_t full_attn_interval = 4; + ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false); + for (uint32_t i = 0; i < hparams.n_layer; ++i) { + hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0); + } + } + + switch (hparams.n_layer) { + case 24: type = LLM_TYPE_2B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; + case LLM_ARCH_QWEN35MOE: + { + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false); + ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true); + + // Load linear attention (gated delta net) parameters + ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); + + // Mark recurrent layers (linear attention layers) + { + uint32_t full_attn_interval = 4; + ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false); + for (uint32_t i = 0; i < hparams.n_layer; ++i) { + hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0); + } + } + + switch (hparams.n_layer) { + case 28: type = LLM_TYPE_35B_A3B; break; + case 48: type = LLM_TYPE_80B_A3B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_MISTRAL3: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -7101,6 +7162,131 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { hparams.n_ff_shexp, n_embd }, 0); } } break; + case LLM_ARCH_QWEN35MOE: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED); + } + + const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used; + + // Calculate dimensions from hyperparameters + const int64_t head_k_dim = hparams.ssm_d_state; + const int64_t head_v_dim = hparams.ssm_d_state; + const int64_t n_k_heads = hparams.ssm_n_group; + const int64_t n_v_heads = hparams.ssm_dt_rank; + const int64_t key_dim = head_k_dim * n_k_heads; + const int64_t value_dim = head_v_dim * n_v_heads; + const int64_t conv_dim = key_dim * 2 + value_dim; + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0); + layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0); + + if (!hparams.is_recurrent(i)) { + // Attention layers + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0); + + // Q/K normalization for attention layers + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0); + } else { + // Linear attention (gated delta net) specific tensors + // Create tensors with calculated dimensions + layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED); + layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), { n_embd, value_dim }, TENSOR_NOT_REQUIRED); + layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), { hparams.ssm_d_conv, conv_dim }, 0); + layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), { hparams.ssm_dt_rank }, 0); + layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, i), { hparams.ssm_dt_rank }, 0); + layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), { n_embd, n_v_heads }, 0); + layer.ssm_alpha = create_tensor(tn(LLM_TENSOR_SSM_ALPHA, "weight", i), { n_embd, n_v_heads }, 0); + layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), { head_v_dim }, 0); + layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), { value_dim, n_embd }, 0); + } + + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0); + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0); + + // Shared experts + const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff; + + layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), { n_embd }, 0); + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, 0); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, 0); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, 0); + } + } break; + case LLM_ARCH_QWEN35: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED); + } + + // Calculate dimensions from hyperparameters + const int64_t head_k_dim = hparams.ssm_d_state; + const int64_t head_v_dim = hparams.ssm_d_state; + const int64_t n_k_heads = hparams.ssm_n_group; + const int64_t n_v_heads = hparams.ssm_dt_rank; + const int64_t key_dim = head_k_dim * n_k_heads; + const int64_t value_dim = head_v_dim * n_v_heads; + const int64_t conv_dim = key_dim * 2 + value_dim; + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0); + layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0); + + if (!hparams.is_recurrent(i)) { + // Attention layers + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0); + + // Q/K normalization for attention layers + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0); + } else { + // Linear attention (gated delta net) specific tensors + // Create tensors with calculated dimensions + layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED); + layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), { n_embd, value_dim }, TENSOR_NOT_REQUIRED); + layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), { hparams.ssm_d_conv, conv_dim }, 0); + layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), { hparams.ssm_dt_rank }, 0); + layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, i), { hparams.ssm_dt_rank }, 0); + layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), { n_embd, n_v_heads }, 0); + layer.ssm_alpha = create_tensor(tn(LLM_TENSOR_SSM_ALPHA, "weight", i), { n_embd, n_v_heads }, 0); + layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), { head_v_dim }, 0); + layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), { value_dim, n_embd }, 0); + } + + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + } + } break; case LLM_ARCH_MIMO2: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -7545,6 +7731,8 @@ void llama_model::print_info() const { arch == LLM_ARCH_PLAMO2 || arch == LLM_ARCH_GRANITE_HYBRID || arch == LLM_ARCH_QWEN3NEXT || + arch == LLM_ARCH_QWEN35 || + arch == LLM_ARCH_QWEN35MOE || arch == LLM_ARCH_NEMOTRON_H || arch == LLM_ARCH_NEMOTRON_H_MOE) { LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv); @@ -8343,6 +8531,14 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { { llm = std::make_unique(*this, params); } break; + case LLM_ARCH_QWEN35: + { + llm = std::make_unique(*this, params); + } break; + case LLM_ARCH_QWEN35MOE: + { + llm = std::make_unique(*this, params); + } break; case LLM_ARCH_MISTRAL3: { llm = std::make_unique(*this, params); @@ -8611,6 +8807,8 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { return LLAMA_ROPE_TYPE_MROPE; case LLM_ARCH_QWEN3VL: case LLM_ARCH_QWEN3VLMOE: + case LLM_ARCH_QWEN35: + case LLM_ARCH_QWEN35MOE: return LLAMA_ROPE_TYPE_IMROPE; case LLM_ARCH_GLM4: diff --git a/src/llama-model.h b/src/llama-model.h index 7b580043b3..adc8ff6479 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -118,6 +118,7 @@ enum llm_type { LLM_TYPE_21B_A3B, // Ernie MoE small LLM_TYPE_30B_A3B, LLM_TYPE_31B_A3_5B, + LLM_TYPE_35B_A3B, // Qwen3.5 LLM_TYPE_48B_A3B, // Kimi Linear LLM_TYPE_80B_A3B, // Qwen3 Next LLM_TYPE_100B_A6B, @@ -322,6 +323,9 @@ struct llama_layer { // qwen3next struct ggml_tensor * ssm_beta_alpha = nullptr; + // qwen3.5 + struct ggml_tensor * ssm_alpha = nullptr; + // rwkv struct ggml_tensor * time_mix_w1 = nullptr; struct ggml_tensor * time_mix_w2 = nullptr; diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index 6d6bdfa090..62e137fb84 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -368,6 +368,13 @@ struct llm_tokenizer_bpe : llm_tokenizer { "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", }; break; + case LLAMA_VOCAB_PRE_TYPE_QWEN35: + regex_exprs = { + // original regex from tokenizer.json + // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; case LLAMA_VOCAB_PRE_TYPE_PORO: case LLAMA_VOCAB_PRE_TYPE_BLOOM: case LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH: @@ -1926,6 +1933,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { tokenizer_pre == "kormo") { pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2; clean_spaces = false; + } else if ( + tokenizer_pre == "qwen35") { + pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN35; + clean_spaces = false; } else if ( tokenizer_pre == "stablelm2") { pre_type = LLAMA_VOCAB_PRE_TYPE_STABLELM2; diff --git a/src/llama-vocab.h b/src/llama-vocab.h index 28c3a82b91..718238fb86 100644 --- a/src/llama-vocab.h +++ b/src/llama-vocab.h @@ -54,6 +54,7 @@ enum llama_vocab_pre_type { LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN = 43, LLAMA_VOCAB_PRE_TYPE_YOUTU = 44, LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE = 45, + LLAMA_VOCAB_PRE_TYPE_QWEN35 = 46, }; struct LLM_KV; diff --git a/src/models/models.h b/src/models/models.h index cfcbb9aaa5..3c66d32531 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -476,6 +476,7 @@ struct llm_build_qwen3vl : public llm_graph_context { struct llm_build_qwen3vlmoe : public llm_graph_context { llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params); }; + struct llm_build_qwen3next : public llm_graph_context_mamba { llm_build_qwen3next(const llama_model & model, const llm_graph_params & params); private: @@ -534,6 +535,124 @@ private: const llama_model & model; }; +struct llm_build_qwen35 : public llm_graph_context_mamba { + llm_build_qwen35(const llama_model & model, const llm_graph_params & params); +private: + ggml_tensor * build_layer_attn( + llm_graph_input_attn_kv * inp_attn, + ggml_tensor * cur, + ggml_tensor * inp_pos, + int * sections, + int il); + + ggml_tensor * build_layer_attn_linear( + llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il); + + ggml_tensor * build_layer_ffn( + ggml_tensor * cur, + int il); + + // returns pair of output and new state + std::pair build_delta_net_chunking( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il); + + // returns pair of output and new state + std::pair build_delta_net_autoregressive( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + int il); + + ggml_tensor * build_norm_gated( + ggml_tensor * input, + ggml_tensor * weights, + ggml_tensor * gate, + int layer); + + // returns pair of qkv, z + std::pair build_qkvz( + ggml_tensor * input, + int il); + + const llama_model & model; +}; + +struct llm_build_qwen35moe : public llm_graph_context_mamba { + llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params); +private: + ggml_tensor * build_layer_attn( + llm_graph_input_attn_kv * inp_attn, + ggml_tensor * cur, + ggml_tensor * inp_pos, + int * sections, + int il); + + ggml_tensor * build_layer_attn_linear( + llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il); + + ggml_tensor * build_layer_ffn( + ggml_tensor * cur, + int il); + + // returns pair of output and new state + std::pair build_delta_net_chunking( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il); + + // returns pair of output and new state + std::pair build_delta_net_autoregressive( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + int il); + + ggml_tensor * build_norm_gated( + ggml_tensor * input, + ggml_tensor * weights, + ggml_tensor * gate, + int layer); + + // returns pair of qkv, z + std::pair build_qkvz( + ggml_tensor * input, + int il); + + const llama_model & model; +}; + struct llm_build_qwen : public llm_graph_context { llm_build_qwen(const llama_model & model, const llm_graph_params & params); }; diff --git a/src/models/qwen35.cpp b/src/models/qwen35.cpp new file mode 100644 index 0000000000..592c170457 --- /dev/null +++ b/src/models/qwen35.cpp @@ -0,0 +1,740 @@ +#include "ggml.h" +#include "models.h" + +#define CHUNK_SIZE 64 + +llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_params & params) : + llm_graph_context_mamba(params), model(model) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + int sections[4]; + std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + cb(inpL, "model.input_embed", -1); + + auto * inp = build_inp_mem_hybrid(); + + ggml_tensor * inp_pos = build_inp_pos(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + ggml_tensor * causal_mask = + ggml_tri(ctx0, ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f), + GGML_TRI_TYPE_LOWER); + + ggml_tensor * identity = ggml_diag(ctx0, ggml_fill(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f)); + ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity); + + ggml_build_forward_expand(gf, causal_mask); + ggml_build_forward_expand(gf, identity); + ggml_build_forward_expand(gf, diag_mask); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // Determine layer type and build appropriate attention mechanism + if (hparams.is_recurrent(il)) { + // Linear attention layer (gated delta net) + cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il); + } else { + // Full attention layer + cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // Residual connection + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "attn_residual", il); + + // Save the tensor before post-attention norm for residual connection + ggml_tensor * ffn_residual = cur; + + // Post-attention norm + ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il); + cb(attn_post_norm, "attn_post_norm", il); + + // Dense FFN layer - without residual connection + cur = build_layer_ffn(attn_post_norm, il); + cb(cur, "ffn_out", il); + + // Residual connection for FFN - add to the tensor from before post_attention_layernorm + cur = ggml_add(ctx0, cur, ffn_residual); + cb(cur, "post_ffn", il); + + // Input for next layer + inpL = cur; + } + cur = inpL; + + // Final norm + cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // LM head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +// utility to get one slice from the third dimension +// input dim: [x, y, c, b] +// output dim: [x, y, 1, b] +static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) { + return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3], + t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c); +} + +std::pair llm_build_qwen35::build_delta_net_chunking( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il) { + const int64_t S_k = q->ne[0]; + const int64_t H_k = q->ne[1]; + const int64_t n_tokens = q->ne[2]; + const int64_t n_seqs = q->ne[3]; + + const int64_t S_v = v->ne[0]; + const int64_t H_v = v->ne[1]; + + GGML_ASSERT(v->ne[2] == n_tokens); + GGML_ASSERT(k->ne[2] == n_tokens); + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); + GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + + GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + + GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + + const float eps_norm = hparams.f_norm_rms_eps; + + q = ggml_l2_norm(ctx0, q, eps_norm); + k = ggml_l2_norm(ctx0, k, eps_norm); + + const float scale = 1.0f / sqrtf(S_v); + + q = ggml_scale(ctx0, q, scale); + + beta = ggml_sigmoid(ctx0, beta); + + cb(q, "q_in", il); + cb(k, "k_in", il); + cb(v, "v_in", il); + cb(beta, "beta_in", il); + cb(g, "g_in", il); + + q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs); + + beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3)); + state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + + cb(q, "q_perm", il); + cb(k, "k_perm", il); + cb(v, "v_perm", il); + cb(beta, "beta_perm", il); + cb(g, "g_perm", il); + cb(state, "state_in", il); + + GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs); + GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs); + GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs); + + // Do padding + const int64_t chunk_size = CHUNK_SIZE; + + const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size; + const int64_t n_chunks = (n_tokens + pad) / chunk_size; + + q = ggml_pad(ctx0, q, 0, pad, 0, 0); + k = ggml_pad(ctx0, k, 0, pad, 0, 0); + v = ggml_pad(ctx0, v, 0, pad, 0, 0); + g = ggml_pad(ctx0, g, pad, 0, 0, 0); + beta = ggml_pad(ctx0, beta, 0, pad, 0, 0); + + cb(q, "q_pad", il); + cb(k, "k_pad", il); + cb(v, "v_pad", il); + cb(beta, "beta_pad", il); + cb(g, "g_pad", il); + + ggml_tensor * v_beta = ggml_mul(ctx0, v, beta); + ggml_tensor * k_beta = ggml_mul(ctx0, k, beta); + + cb(v_beta, "v_beta", il); + cb(k_beta, "k_beta", il); + + q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs); + k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs); + k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs); + v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs); + v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs); + + g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs); + beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs); + + ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g); + cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs); + ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs); + + ggml_tensor * gcs_j_broadcast = + ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs); + + ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i); + cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); + decay_mask = ggml_exp(ctx0, decay_mask); + decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); + + ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta); + + ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask); + ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask)); + cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask); + ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower); + + ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false); + attn = ggml_mul(ctx0, lin_solve, causal_mask); + attn = ggml_add(ctx0, attn, identity); + cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn); + + ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum)); + ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t); + + ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp); + cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * k_cumdecay = + ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp))))); + cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q); + attn_kq = ggml_mul(ctx0, attn_kq, decay_mask); + attn_kq = ggml_mul(ctx0, attn_kq, diag_mask); + cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + + // vectorized calculation of key_gdiff + // improved from the chunked version: + // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1) + // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp() + // key_gdiff = key * g_diff.unsqueeze(-1) + // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new + // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew + + // get last element in g_cumsum along chunk_size dimension (ne0) + // example: [[x, y, z, ..., last], ...] -> [[last], ...] + ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3], + g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3], + (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum)); + g_last = ggml_cont(ctx0, g_last); + cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last); + cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last)); + cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff); + ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp, + 1, chunk_size, n_chunks, g_diff_exp->ne[3]); + + ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t); + cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)); + cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs) + + // state to be updated per chunk + ggml_tensor * new_state = state; // ggml_dup(ctx0, state); + cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs) + + // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs) + ggml_tensor * core_attn_out = nullptr; + + for (int64_t chunk = 0; chunk < n_chunks; chunk++) { + // shape: (S_k, chunk_size, 1, H_k * n_seqs) + ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul + + // shape: (S_v, chunk_size, 1, H_v * n_seqs) + ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat + + // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul + + // shape: (chunk_size, 1, H_v * n_seqs) + ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat + + // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0) + // replaced by precomputed attn_kq + ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk); + cb(attn_chunk, "attn_chunk", il); + + ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs); + + // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state + ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk); + cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs) + + // v_new = v_i - v_prime + ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime); + ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new)); + cb(v_new, "v_new_chunk", il); + + // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state + ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk); + ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp); + cb(attn_inter, "attn_inter_chunk", il); + + // core_attn_out[:, :, i] = attn_inter + attn @ v_new + ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk); + cb(v_attn, "v_attn_chunk", il); + + ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn); + cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs) + + core_attn_out = core_attn_out == nullptr + ? core_attn_out_chunk + : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2); + + // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new + ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk); + //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why? + ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t); + + // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew + ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk)); + new_state = ggml_add(ctx0, + ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)), + ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs)); + } + + // truncate padded tokens + ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out, + S_v, n_tokens, H_v, n_seqs, + ggml_row_size(core_attn_out->type, S_v), + ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks), + ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0); + output_tokens = ggml_cont(ctx0, output_tokens); + cb(output_tokens, "output_tokens", il); + + // permute back to (S_v, H_v, n_tokens, n_seqs) + output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3); + output_tokens = ggml_cont(ctx0, output_tokens); + + return {output_tokens, new_state}; +} + +std::pair llm_build_qwen35::build_delta_net_autoregressive( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + int il) { + const int64_t S_k = q->ne[0]; + const int64_t H_k = q->ne[1]; + const int64_t n_tokens = q->ne[2]; + const int64_t n_seqs = q->ne[3]; + + const int64_t S_v = v->ne[0]; + const int64_t H_v = v->ne[1]; + + GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing + GGML_ASSERT(v->ne[2] == n_tokens); + GGML_ASSERT(k->ne[2] == n_tokens); + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); + GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + + GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + + GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + + const float eps_norm = hparams.f_norm_rms_eps; + + q = ggml_l2_norm(ctx0, q, eps_norm); + k = ggml_l2_norm(ctx0, k, eps_norm); + + const float scale = 1.0f / sqrtf(S_v); + + q = ggml_scale(ctx0, q, scale); + beta = ggml_sigmoid(ctx0, beta); + + cb(q, "q_in", il); + cb(k, "k_in", il); + cb(v, "v_in", il); + cb(beta, "beta_in", il); + cb(g, "g_in", il); + + state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + + ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs); + ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs); + + // Apply exponential to g_t + g_t = ggml_exp(ctx0, g_t); + + // Apply the gated delta rule for the single timestep + // last_recurrent_state = last_recurrent_state * g_t + state = ggml_mul(ctx0, state, g_t); + + // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2) + ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs); + ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed); + // we need to sum over dim=-2, so we transpose, sum, then transpose again + kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem)))); + + // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v) + ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs); + // delta = (v_t - kv_mem) * beta_t + ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs] + ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t); + + // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta + ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta); + state = ggml_add(ctx0, state, k_t_delta); + + // Compute the attention output + // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2) + ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t + ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed); + // again, since it's over dim = -2, transpose, sum, transpose back + ggml_tensor * core_attn_out = + ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q)))); + + // core_attn_out should be [S_v, 1, H_v, n_seqs] after this + cb(core_attn_out, "output_tokens", il); + cb(state, "new_state", il); + + return {core_attn_out, state}; +} + +std::pair llm_build_qwen35::build_qkvz( + ggml_tensor * input, + int il) { + const int64_t n_seqs = ubatch.n_seqs; + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input); + qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs); + cb(qkv_mixed, "linear_attn_qkv_mixed", il); + + ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input); + cb(z, "z", il); + + return { qkv_mixed, z }; +} + +ggml_tensor * llm_build_qwen35::build_norm_gated( + ggml_tensor * input, + ggml_tensor * weights, + ggml_tensor * gate, + int layer) { + ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer); + ggml_tensor * gated_silu = ggml_silu(ctx0, gate); + + return ggml_mul(ctx0, normalized, gated_silu); +} + +ggml_tensor * llm_build_qwen35::build_layer_attn( + llm_graph_input_attn_kv * inp, + ggml_tensor * cur, + ggml_tensor * inp_pos, + int * sections, + int il) { + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention + + // Qwen3Next uses a single Q projection that outputs query + gate + ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur); // [ (n_embd_head * 2) * n_head, n_tokens ] + cb(Qcur_full, "Qcur_full", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur_full) * n_embd_head * 2, + ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, 0); + cb(Qcur, "Qcur_reshaped", il); + + // Apply Q normalization + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + // Apply K normalization + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur_full) * n_embd_head * 2, + ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, + ggml_element_size(Qcur_full) * n_embd_head); + gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens); + cb(gate, "gate_reshaped", il); + + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + // Apply MRoPE + Qcur = ggml_rope_multi( + ctx0, Qcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_multi( + ctx0, Kcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // Attention computation + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + cur = build_attn(inp, + nullptr, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_pregate", il); + + ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate); + cb(gate_sigmoid, "gate_sigmoid", il); + + cur = ggml_mul(ctx0, cur, gate_sigmoid); + cb(cur, "attn_gated", il); + + cur = build_lora_mm(model.layers[il].wo, cur); + cb(cur, "attn_output", il); + + return cur; +} + +ggml_tensor * llm_build_qwen35::build_layer_attn_linear( + llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il) { + const auto * mctx_cur = inp->mctx; + + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t n_seqs = ubatch.n_seqs; + const int64_t head_k_dim = hparams.ssm_d_state; + const int64_t num_k_heads = hparams.ssm_n_group; + const int64_t num_v_heads = hparams.ssm_dt_rank; + const int64_t head_v_dim = d_inner / num_v_heads; + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + const auto kv_head = mctx_cur->get_head(); + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs()); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + // Input projections + auto qkvz = build_qkvz(cur, il); + ggml_tensor * qkv_mixed = qkvz.first; + ggml_tensor * z = qkvz.second; + + ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur); + beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs); + cb(beta, "beta", il); + ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur); + alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs); + cb(alpha, "alpha", il); + + ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt); + ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased); + cb(alpha_softplus, "a_softplus", il); + ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus + cb(gate, "gate", il); + + // Get convolution states from cache + ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); + ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); + + // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state(); + + // Build the convolution states tensor + ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); + cb(conv_states, "conv_states", il); + + // Calculate convolution kernel size + ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d; + const int64_t conv_kernel_size = conv_kernel->ne[0]; + const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state; + conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs); + cb(conv_states, "conv_states_reshaped", il); + + qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3); + cb(qkv_mixed, "qkv_mixed_permuted", il); + + ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0); + cb(conv_input, "conv_input", il); + + // Update convolution state cache + // Extract the last (conv_kernel_size - 1) states from conv_input + ggml_tensor * last_conv_states = + ggml_view_3d(ctx0, conv_input, conv_kernel_size - 1, conv_channels, n_seqs, conv_input->nb[1], + conv_input->nb[2], (conv_input->ne[0] - conv_states->ne[0]) * ggml_element_size(conv_input)); + cb(last_conv_states, "last_conv_states", il); + + ggml_tensor * state_update_target = + ggml_view_1d(ctx0, conv_states_all, (conv_kernel_size - 1) * conv_channels * n_seqs, + kv_head * (conv_kernel_size - 1) * conv_channels * ggml_element_size(conv_states_all)); + cb(state_update_target, "state_update_target", il); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target)); + cb(conv_states_all, "conv_states_updated", il); + + // Apply SSM convolution + ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel); + cb(conv_output_proper, "conv_output_raw", il); + + ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper); + cb(conv_output_silu, "conv_output_silu", il); + + ggml_tensor * conv_qkv_mix = conv_output_silu; + + // Calculate the total conv dimension + int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads; + int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim); + + // Extract the convolved Q, K, V from conv_output + ggml_tensor * q_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0); + cb(q_conv, "q_conv", il); + ggml_tensor * k_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, + head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + cb(k_conv, "k_conv", il); + ggml_tensor * v_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv, + 2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + cb(v_conv, "v_conv", il); + + // Unsqueeze them + q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); + + ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs); + state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs); + cb(state, "state_predelta", il); + + // if head keys and value keys are different, repeat Q/K to match V's head count + // V heads are in tiled order (from conversion), so simple tiled repeat works + if (num_k_heads != num_v_heads) { + GGML_ASSERT(num_v_heads % num_k_heads == 0); + q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs); + k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs); + } + + cb(q_conv, "q_conv_predelta", il); + cb(k_conv, "k_conv_predelta", il); + cb(v_conv, "v_conv_predelta", il); + + // Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens + std::pair attn_out; // pair of (output, new_state) + if (n_seq_tokens == 1) { + attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il); + } else { + attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il); + } + ggml_tensor * output = attn_out.first; + ggml_tensor * new_state = attn_out.second; + cb(output, "attn_output", il); + cb(new_state, "new_state", il); + + // Update the recurrent states + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, new_state, + ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs, + kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all)))); + + // Reshape both attn_out_final and z to 2D tensors for normalization + // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] + ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + + // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] + ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + + // Apply gated normalization: self.norm(core_attn_out, z) + ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il); + + // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim] + ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs); + cb(final_output, "final_output", il); + + // Output projection + cur = build_lora_mm(model.layers[il].ssm_out, final_output); + cb(cur, "linear_attn_out", il); + + // Reshape back to original dimensions + cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs); + return cur; +} + +ggml_tensor * llm_build_qwen35::build_layer_ffn(ggml_tensor * cur, const int il) { + // Qwen3.5 does not use MoE FFN + GGML_ASSERT(model.layers[il].ffn_gate_inp == nullptr); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + return cur; +} diff --git a/src/models/qwen35moe.cpp b/src/models/qwen35moe.cpp new file mode 100644 index 0000000000..0db8f825c6 --- /dev/null +++ b/src/models/qwen35moe.cpp @@ -0,0 +1,774 @@ +#include "ggml.h" +#include "models.h" + +#define CHUNK_SIZE 64 + +llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params) : + llm_graph_context_mamba(params), model(model) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + int sections[4]; + std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + cb(inpL, "model.input_embed", -1); + + auto * inp = build_inp_mem_hybrid(); + + ggml_tensor * inp_pos = build_inp_pos(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + ggml_tensor * causal_mask = + ggml_tri(ctx0, ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f), + GGML_TRI_TYPE_LOWER); + + ggml_tensor * identity = ggml_diag(ctx0, ggml_fill(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f)); + ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity); + + ggml_build_forward_expand(gf, causal_mask); + ggml_build_forward_expand(gf, identity); + ggml_build_forward_expand(gf, diag_mask); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // Determine layer type and build appropriate attention mechanism + if (hparams.is_recurrent(il)) { + // Linear attention layer (gated delta net) + cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il); + } else { + // Full attention layer + cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // Residual connection + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "attn_residual", il); + + // Save the tensor before post-attention norm for residual connection + ggml_tensor * ffn_residual = cur; + + // Post-attention norm + ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il); + cb(attn_post_norm, "attn_post_norm", il); + + // MOE FFN layer + cur = build_layer_ffn(attn_post_norm, il); + cb(cur, "ffn_out", il); + + // Residual connection for FFN - add to the tensor from before post_attention_layernorm + cur = ggml_add(ctx0, cur, ffn_residual); + cb(cur, "post_moe", il); + + // Input for next layer + inpL = cur; + } + cur = inpL; + + // Final norm + cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // LM head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +// utility to get one slice from the third dimension +// input dim: [x, y, c, b] +// output dim: [x, y, 1, b] +static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) { + return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3], + t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c); +} + +std::pair llm_build_qwen35moe::build_delta_net_chunking( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il) { + const int64_t S_k = q->ne[0]; + const int64_t H_k = q->ne[1]; + const int64_t n_tokens = q->ne[2]; + const int64_t n_seqs = q->ne[3]; + + const int64_t S_v = v->ne[0]; + const int64_t H_v = v->ne[1]; + + GGML_ASSERT(v->ne[2] == n_tokens); + GGML_ASSERT(k->ne[2] == n_tokens); + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); + GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + + GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + + GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + + const float eps_norm = hparams.f_norm_rms_eps; + + q = ggml_l2_norm(ctx0, q, eps_norm); + k = ggml_l2_norm(ctx0, k, eps_norm); + + const float scale = 1.0f / sqrtf(S_v); + + q = ggml_scale(ctx0, q, scale); + + beta = ggml_sigmoid(ctx0, beta); + + cb(q, "q_in", il); + cb(k, "k_in", il); + cb(v, "v_in", il); + cb(beta, "beta_in", il); + cb(g, "g_in", il); + + q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs); + + beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3)); + state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + + cb(q, "q_perm", il); + cb(k, "k_perm", il); + cb(v, "v_perm", il); + cb(beta, "beta_perm", il); + cb(g, "g_perm", il); + cb(state, "state_in", il); + + GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs); + GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs); + GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs); + + // Do padding + const int64_t chunk_size = CHUNK_SIZE; + + const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size; + const int64_t n_chunks = (n_tokens + pad) / chunk_size; + + q = ggml_pad(ctx0, q, 0, pad, 0, 0); + k = ggml_pad(ctx0, k, 0, pad, 0, 0); + v = ggml_pad(ctx0, v, 0, pad, 0, 0); + g = ggml_pad(ctx0, g, pad, 0, 0, 0); + beta = ggml_pad(ctx0, beta, 0, pad, 0, 0); + + cb(q, "q_pad", il); + cb(k, "k_pad", il); + cb(v, "v_pad", il); + cb(beta, "beta_pad", il); + cb(g, "g_pad", il); + + ggml_tensor * v_beta = ggml_mul(ctx0, v, beta); + ggml_tensor * k_beta = ggml_mul(ctx0, k, beta); + + cb(v_beta, "v_beta", il); + cb(k_beta, "k_beta", il); + + q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs); + k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs); + k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs); + v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs); + v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs); + + g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs); + beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs); + + ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g); + cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs); + ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs); + + ggml_tensor * gcs_j_broadcast = + ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs); + + ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i); + cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); + decay_mask = ggml_exp(ctx0, decay_mask); + decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); + + ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta); + + ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask); + ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask)); + cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask); + ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower); + + ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false); + attn = ggml_mul(ctx0, lin_solve, causal_mask); + attn = ggml_add(ctx0, attn, identity); + cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn); + + ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum)); + ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t); + + ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp); + cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * k_cumdecay = + ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp))))); + cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q); + attn_kq = ggml_mul(ctx0, attn_kq, decay_mask); + attn_kq = ggml_mul(ctx0, attn_kq, diag_mask); + cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + + // vectorized calculation of key_gdiff + // improved from the chunked version: + // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1) + // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp() + // key_gdiff = key * g_diff.unsqueeze(-1) + // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new + // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew + + // get last element in g_cumsum along chunk_size dimension (ne0) + // example: [[x, y, z, ..., last], ...] -> [[last], ...] + ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3], + g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3], + (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum)); + g_last = ggml_cont(ctx0, g_last); + cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last); + cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last)); + cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff); + ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp, + 1, chunk_size, n_chunks, g_diff_exp->ne[3]); + + ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t); + cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)); + cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs) + + + // state to be updated per chunk + ggml_tensor * new_state = state; // ggml_dup(ctx0, state); + cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs) + + // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs) + ggml_tensor * core_attn_out = nullptr; + + for (int64_t chunk = 0; chunk < n_chunks; chunk++) { + // shape: (S_k, chunk_size, 1, H_k * n_seqs) + ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul + + // shape: (S_v, chunk_size, 1, H_v * n_seqs) + ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat + + // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul + + // shape: (chunk_size, 1, H_v * n_seqs) + ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat + + // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0) + // replaced by precomputed attn_kq + ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk); + cb(attn_chunk, "attn_chunk", il); + + ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs); + + // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state + ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk); + cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs) + + // v_new = v_i - v_prime + ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime); + ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new)); + cb(v_new, "v_new_chunk", il); + + // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state + ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk); + ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp); + cb(attn_inter, "attn_inter_chunk", il); + + // core_attn_out[:, :, i] = attn_inter + attn @ v_new + ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk); + cb(v_attn, "v_attn_chunk", il); + + ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn); + cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs) + + core_attn_out = core_attn_out == nullptr + ? core_attn_out_chunk + : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2); + + // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new + ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk); + //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why? + ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t); + + // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew + ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk)); + new_state = ggml_add(ctx0, + ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)), + ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs)); + } + + // truncate padded tokens + ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out, + S_v, n_tokens, H_v, n_seqs, + ggml_row_size(core_attn_out->type, S_v), + ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks), + ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0); + output_tokens = ggml_cont(ctx0, output_tokens); + cb(output_tokens, "output_tokens", il); + + // permute back to (S_v, H_v, n_tokens, n_seqs) + output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3); + output_tokens = ggml_cont(ctx0, output_tokens); + + return {output_tokens, new_state}; +} + +std::pair llm_build_qwen35moe::build_delta_net_autoregressive( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + int il) { + const int64_t S_k = q->ne[0]; + const int64_t H_k = q->ne[1]; + const int64_t n_tokens = q->ne[2]; + const int64_t n_seqs = q->ne[3]; + + const int64_t S_v = v->ne[0]; + const int64_t H_v = v->ne[1]; + + GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing + GGML_ASSERT(v->ne[2] == n_tokens); + GGML_ASSERT(k->ne[2] == n_tokens); + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); + GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + + GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + + GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + + const float eps_norm = hparams.f_norm_rms_eps; + + q = ggml_l2_norm(ctx0, q, eps_norm); + k = ggml_l2_norm(ctx0, k, eps_norm); + + const float scale = 1.0f / sqrtf(S_v); + + q = ggml_scale(ctx0, q, scale); + beta = ggml_sigmoid(ctx0, beta); + + cb(q, "q_in", il); + cb(k, "k_in", il); + cb(v, "v_in", il); + cb(beta, "beta_in", il); + cb(g, "g_in", il); + + state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + + ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs); + ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs); + + // Apply exponential to g_t + g_t = ggml_exp(ctx0, g_t); + + // Apply the gated delta rule for the single timestep + // last_recurrent_state = last_recurrent_state * g_t + state = ggml_mul(ctx0, state, g_t); + + // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2) + ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs); + ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed); + // we need to sum over dim=-2, so we transpose, sum, then transpose again + kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem)))); + + // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v) + ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs); + // delta = (v_t - kv_mem) * beta_t + ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs] + ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t); + + // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta + ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta); + state = ggml_add(ctx0, state, k_t_delta); + + // Compute the attention output + // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2) + ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t + ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed); + // again, since it's over dim = -2, transpose, sum, transpose back + ggml_tensor * core_attn_out = + ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q)))); + + // core_attn_out should be [S_v, 1, H_v, n_seqs] after this + cb(core_attn_out, "output_tokens", il); + cb(state, "new_state", il); + + return {core_attn_out, state}; +} + +std::pair llm_build_qwen35moe::build_qkvz( + ggml_tensor * input, + int il) { + const int64_t n_seqs = ubatch.n_seqs; + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input); + qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs); + cb(qkv_mixed, "linear_attn_qkv_mixed", il); + + ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input); + cb(z, "z", il); + + return { qkv_mixed, z }; +} + +ggml_tensor * llm_build_qwen35moe::build_norm_gated( + ggml_tensor * input, + ggml_tensor * weights, + ggml_tensor * gate, + int layer) { + ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer); + ggml_tensor * gated_silu = ggml_silu(ctx0, gate); + + return ggml_mul(ctx0, normalized, gated_silu); +} + +ggml_tensor * llm_build_qwen35moe ::build_layer_attn( + llm_graph_input_attn_kv * inp, + ggml_tensor * cur, + ggml_tensor * inp_pos, + int * sections, + int il) { + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention + + // Qwen3Next uses a single Q projection that outputs query + gate + ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur); // [ (n_embd_head * 2) * n_head, n_tokens ] + cb(Qcur_full, "Qcur_full", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur_full) * n_embd_head * 2, + ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, 0); + cb(Qcur, "Qcur_reshaped", il); + + // Apply Q normalization + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + // Apply K normalization + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur_full) * n_embd_head * 2, + ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, + ggml_element_size(Qcur_full) * n_embd_head); + gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens); + cb(gate, "gate_reshaped", il); + + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + // Apply IMRoPE + Qcur = ggml_rope_multi( + ctx0, Qcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_multi( + ctx0, Kcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // Attention computation + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + cur = build_attn(inp, + nullptr, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_pregate", il); + + ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate); + cb(gate_sigmoid, "gate_sigmoid", il); + + cur = ggml_mul(ctx0, cur, gate_sigmoid); + cb(cur, "attn_gated", il); + + cur = build_lora_mm(model.layers[il].wo, cur); + cb(cur, "attn_output", il); + + return cur; +} + +ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear( + llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il) { + const auto * mctx_cur = inp->mctx; + + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t n_seqs = ubatch.n_seqs; + const int64_t head_k_dim = hparams.ssm_d_state; + const int64_t num_k_heads = hparams.ssm_n_group; + const int64_t num_v_heads = hparams.ssm_dt_rank; + const int64_t head_v_dim = d_inner / num_v_heads; + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + const auto kv_head = mctx_cur->get_head(); + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs()); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + // Input projections + auto qkvz = build_qkvz(cur, il); + ggml_tensor * qkv_mixed = qkvz.first; + ggml_tensor * z = qkvz.second; + + ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur); + beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs); + cb(beta, "beta", il); + ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur); + alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs); + cb(alpha, "alpha", il); + + ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt); + ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased); + cb(alpha_softplus, "a_softplus", il); + ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus + cb(gate, "gate", il); + + // Get convolution states from cache + ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); + ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); + + // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state(); + + // Build the convolution states tensor + ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); + cb(conv_states, "conv_states", il); + + // Calculate convolution kernel size + ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d; + const int64_t conv_kernel_size = conv_kernel->ne[0]; + const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state; + conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs); + cb(conv_states, "conv_states_reshaped", il); + + qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3); + cb(qkv_mixed, "qkv_mixed_permuted", il); + + ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0); + cb(conv_input, "conv_input", il); + + // Update convolution state cache + // Extract the last (conv_kernel_size - 1) states from conv_input + ggml_tensor * last_conv_states = + ggml_view_3d(ctx0, conv_input, conv_kernel_size - 1, conv_channels, n_seqs, conv_input->nb[1], + conv_input->nb[2], (conv_input->ne[0] - conv_states->ne[0]) * ggml_element_size(conv_input)); + cb(last_conv_states, "last_conv_states", il); + + ggml_tensor * state_update_target = + ggml_view_1d(ctx0, conv_states_all, (conv_kernel_size - 1) * conv_channels * n_seqs, + kv_head * (conv_kernel_size - 1) * conv_channels * ggml_element_size(conv_states_all)); + cb(state_update_target, "state_update_target", il); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target)); + cb(conv_states_all, "conv_states_updated", il); + + // Apply SSM convolution + ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel); + cb(conv_output_proper, "conv_output_raw", il); + + ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper); + cb(conv_output_silu, "conv_output_silu", il); + + ggml_tensor * conv_qkv_mix = conv_output_silu; + + // Calculate the total conv dimension + int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads; + int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim); + + // Extract the convolved Q, K, V from conv_output + ggml_tensor * q_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0); + cb(q_conv, "q_conv", il); + ggml_tensor * k_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, + head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + cb(k_conv, "k_conv", il); + ggml_tensor * v_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv, + 2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + cb(v_conv, "v_conv", il); + + // Unsqueeze them + q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); + + ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs); + state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs); + cb(state, "state_predelta", il); + + // if head keys and value keys are different, repeat Q/K to match V's head count + // V heads are in tiled order (from conversion), so simple tiled repeat works + if (num_k_heads != num_v_heads) { + GGML_ASSERT(num_v_heads % num_k_heads == 0); + q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs); + k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs); + } + + cb(q_conv, "q_conv_predelta", il); + cb(k_conv, "k_conv_predelta", il); + cb(v_conv, "v_conv_predelta", il); + + // Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens + std::pair attn_out; // pair of (output, new_state) + if (n_seq_tokens == 1) { + attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il); + } else { + attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il); + } + ggml_tensor * output = attn_out.first; + ggml_tensor * new_state = attn_out.second; + cb(output, "attn_output", il); + cb(new_state, "new_state", il); + + // Update the recurrent states + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, new_state, + ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs, + kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all)))); + + // Reshape both attn_out_final and z to 2D tensors for normalization + // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] + ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + + // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] + ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + + // Apply gated normalization: self.norm(core_attn_out, z) + ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il); + + // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim] + ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs); + cb(final_output, "final_output", il); + + // Output projection + cur = build_lora_mm(model.layers[il].ssm_out, final_output); + cb(cur, "linear_attn_out", il); + + // Reshape back to original dimensions + cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs); + return cur; +} + +ggml_tensor * llm_build_qwen35moe ::build_layer_ffn(ggml_tensor * cur, const int il) { + // Check if this is an MoE layer + GGML_ASSERT(model.layers[il].ffn_gate_inp != nullptr); + + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, LLM_FFN_SILU, + true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il); + cb(moe_out, "ffn_moe_out", il); + + // Add shared experts if present - following Qwen3Next reference implementation + if (model.layers[il].ffn_up_shexp != nullptr) { + ggml_tensor * ffn_shexp = + build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + // Apply shared expert gating as in the reference implementation + // The shared expert has its own gate that is sigmoided + // Note: ffn_gate_inp_shexp is the shared expert gate (outputs 1 value per token) + ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur); + cb(shared_gate, "shared_expert_gate", il); + + // Apply sigmoid to the gate + shared_gate = ggml_sigmoid(ctx0, shared_gate); + cb(shared_gate, "shared_expert_gate_sigmoid", il); + + + // Apply the gate to the shared expert output + ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate); + cb(ffn_shexp, "ffn_shexp_gated", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } else { + cur = moe_out; + } + + return cur; +} diff --git a/tools/mtmd/models/qwen3vl.cpp b/tools/mtmd/models/qwen3vl.cpp index 35a42cb84d..5ecb10fe43 100644 --- a/tools/mtmd/models/qwen3vl.cpp +++ b/tools/mtmd/models/qwen3vl.cpp @@ -182,7 +182,9 @@ ggml_cgraph * clip_graph_qwen3vl::build() { model.mm_1_w, model.mm_1_b, ffn_op_type::FFN_GELU, -1); - embeddings = ggml_concat(ctx0, embeddings, deepstack_features, 0); // concat along the feature dimension + if (deepstack_features) { + embeddings = ggml_concat(ctx0, embeddings, deepstack_features, 0); + } // concat along the feature dimension // build the graph ggml_build_forward_expand(gf, embeddings);