diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index c167de8a46..843c00a896 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -920,7 +920,7 @@ class TextModel(ModelBase): self.gguf_writer.add_expert_group_used_count(n_group_used) logger.info(f"gguf: expert groups used count = {n_group_used}") - if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation_func"], optional=True)) is not None: + if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation", "moe_router_activation_func"], optional=True)) is not None: if score_func == "sigmoid": self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID) elif score_func == "softmax": @@ -7912,6 +7912,135 @@ class MimoV2Model(TextModel): raise ValueError(f"Unprocessed experts: {experts}") +@ModelBase.register("Step3p5ForCausalLM") +class Step35Model(TextModel): + model_arch = gguf.MODEL_ARCH.STEP35 + + def set_gguf_parameters(self): + rope_theta = self.hparams.get("rope_theta") + if isinstance(rope_theta, list): + self.hparams["rope_theta"] = float(rope_theta[0]) + self.hparams["local_rope_theta"] = float(rope_theta[1]) + self.rope_parameters["rope_theta"] = self.hparams["rope_theta"] + self.rope_parameters["sliding_attention"] = {"rope_theta": self.hparams["local_rope_theta"]} + + super().set_gguf_parameters() + + layer_types = self.hparams.get("layer_types") or [] + partial_rotary_factors = self.hparams.get("partial_rotary_factors") or [] + attn_other = self.hparams.get("attention_other_setting") or {} + + n_head_base = self.hparams["num_attention_heads"] + n_kv_base = self.hparams["num_attention_groups"] + + n_head_swa = attn_other.get("num_attention_heads", n_head_base) + n_kv_swa = attn_other.get("num_attention_groups", n_kv_base) + + layer_types = layer_types[: self.block_count] + partial_rotary_factors = partial_rotary_factors[: self.block_count] + assert [1.0 if lt == "sliding_attention" else 0.5 for lt in layer_types] == partial_rotary_factors + head_arr = [n_head_swa if lt == "sliding_attention" else n_head_base for lt in layer_types] + kv_arr = [n_kv_swa if lt == "sliding_attention" else n_kv_base for lt in layer_types] + swa_pat = [lt == "sliding_attention" for lt in layer_types] + + self.gguf_writer.add_head_count(head_arr) + self.gguf_writer.add_head_count_kv(kv_arr) + + self.gguf_writer.add_sliding_window(self.hparams["sliding_window"]) + self.gguf_writer.add_sliding_window_pattern(swa_pat) + + self.gguf_writer.add_value_length(self.hparams["head_dim"]) + + # MoE params + self.gguf_writer.add_expert_count(self.hparams["moe_num_experts"]) + self.gguf_writer.add_expert_used_count(self.hparams["moe_top_k"]) + self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"]) + self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["share_expert_dim"]) + + if (moe_router_scaling_factor := self.hparams.get("moe_router_scaling_factor")) is not None: + self.gguf_writer.add_expert_weights_scale(moe_router_scaling_factor) + if (norm_expert_weight := self.hparams.get("norm_expert_weight")) is not None: + self.gguf_writer.add_expert_weights_norm(norm_expert_weight) + + # leading dense blocks + leading_dense = 0 + moe_layers_enum = self.hparams.get("moe_layers_enum") + if isinstance(moe_layers_enum, str) and moe_layers_enum.strip(): + moe_layers = sorted(int(i) for i in moe_layers_enum.strip().split(",")) + if moe_layers: + leading_dense = max(0, moe_layers[0]) + self.gguf_writer.add_leading_dense_block_count(leading_dense) + self.gguf_writer.add_moe_every_n_layers(int(self.hparams.get("moe_every_n_layer", 1))) + + self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-5)) + + # Optional per-layer SwiGLU clamps. + if (limits := self.hparams.get("swiglu_limits")) is not None: + limits_f = [0.0 if v is None else float(v) for v in limits[: self.block_count]] + self.gguf_writer.add_swiglu_clamp_exp(limits_f) + if (limits_shared := self.hparams.get("swiglu_limits_shared")) is not None: + limits_shared_f = [0.0 if v is None else float(v) for v in limits_shared[: self.block_count]] + self.gguf_writer.add_swiglu_clamp_shexp(limits_shared_f) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None): + # remove mtp layers + if (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None: + il = int(m.group(1)) + n_main = int(self.hparams.get("num_hidden_layers", self.block_count)) + if il >= n_main: + return + if name.endswith("norm.weight"): + data_torch += 1.0 + # Map router bias (expert selection bias) to a GGUF bias tensor + if name.endswith(".moe.router_bias"): + name += ".bias" + + if name.endswith((".self_attn.g_proj.weight", ".moe.gate.weight", ".moe.up_proj.weight", ".moe.gate_proj.weight", ".moe.down_proj.weight")): + data_torch = data_torch.squeeze().contiguous() + + yield from super().modify_tensors(data_torch, name, bid) + + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: + # Step35 can optionally use Llama-3 style RoPE scaling (HF: rope_scaling.rope_type == "llama3"). + # llama.cpp represents this via a single extra tensor: "rope_freqs.weight" (aka MODEL_TENSOR.ROPE_FREQS). + rope_params = self.rope_parameters.get("full_attention", self.rope_parameters) + rope_type = rope_params.get("rope_type") or "" + if rope_type.lower() != "llama3": + return + + # Step35 configs can carry per-layer rope_theta as a list; for llama3 rope factors we use the base value. + rope_theta = self.hparams.get("rope_theta", 10000.0) + if isinstance(rope_theta, list): + rope_theta = rope_theta[0] + base = float(rope_theta) + if (dim := self.hparams.get("head_dim")) is None: + dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"] + dim = int(dim) + + freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + + factor = float(rope_params.get("factor", 8.0)) + low_freq_factor = float(rope_params.get("low_freq_factor", 1.0)) + high_freq_factor = float(rope_params.get("high_freq_factor", 4.0)) + old_context_len = int(rope_params.get("original_max_position_embeddings", self.hparams.get("original_max_position_embeddings", 8192))) + + low_freq_wavelen = old_context_len / low_freq_factor + high_freq_wavelen = old_context_len / high_freq_factor + + rope_factors: list[float] = [] + for freq in freqs: + wavelen = 2 * math.pi / float(freq) + if wavelen < high_freq_wavelen: + rope_factors.append(1.0) + elif wavelen > low_freq_wavelen: + rope_factors.append(factor) + else: + smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor) + rope_factors.append(1.0 / ((1.0 - smooth) / factor + smooth)) + + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32)) + + @ModelBase.register("PanguEmbeddedForCausalLM") class PanguEmbeddedModel(TextModel): model_arch = gguf.MODEL_ARCH.PANGU_EMBED diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 3ddbc73d1c..3af4fffe95 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -146,6 +146,8 @@ class Keys: ALTUP_ACTIVE_IDX = "{arch}.altup.active_idx" ALTUP_NUM_INPUTS = "{arch}.altup.num_inputs" EMBD_LENGTH_PER_LAYER_INP = "{arch}.embedding_length_per_layer_input" + SWIGLU_CLAMP_EXP = "{arch}.swiglu_clamp_exp" + SWIGLU_CLAMP_SHEXP = "{arch}.swiglu_clamp_shexp" DENSE_FEAT_IN_SIZE = "{arch}.{dense}_feat_in" DENSE_FEAT_OUT_SIZE = "{arch}.{dense}_feat_out" @@ -179,20 +181,20 @@ class Keys: TEMPERATURE_SCALE = "{arch}.attention.temperature_scale" class Rope: - DIMENSION_COUNT = "{arch}.rope.dimension_count" - DIMENSION_SECTIONS = "{arch}.rope.dimension_sections" - FREQ_BASE = "{arch}.rope.freq_base" - FREQ_BASE_SWA = "{arch}.rope.freq_base_swa" - SCALING_TYPE = "{arch}.rope.scaling.type" - SCALING_FACTOR = "{arch}.rope.scaling.factor" - SCALING_ATTN_FACTOR = "{arch}.rope.scaling.attn_factor" - SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length" - SCALING_FINETUNED = "{arch}.rope.scaling.finetuned" - SCALING_YARN_LOG_MUL = "{arch}.rope.scaling.yarn_log_multiplier" - SCALING_YARN_EXT_FACTOR = "{arch}.rope.scaling.yarn_ext_factor" - SCALING_YARN_ATTN_FACTOR = "{arch}.rope.scaling.yarn_attn_factor" - SCALING_YARN_BETA_FAST = "{arch}.rope.scaling.yarn_beta_fast" - SCALING_YARN_BETA_SLOW = "{arch}.rope.scaling.yarn_beta_slow" + DIMENSION_COUNT = "{arch}.rope.dimension_count" + DIMENSION_SECTIONS = "{arch}.rope.dimension_sections" + FREQ_BASE = "{arch}.rope.freq_base" + FREQ_BASE_SWA = "{arch}.rope.freq_base_swa" + SCALING_TYPE = "{arch}.rope.scaling.type" + SCALING_FACTOR = "{arch}.rope.scaling.factor" + SCALING_ATTN_FACTOR = "{arch}.rope.scaling.attn_factor" + SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length" + SCALING_FINETUNED = "{arch}.rope.scaling.finetuned" + SCALING_YARN_LOG_MUL = "{arch}.rope.scaling.yarn_log_multiplier" + SCALING_YARN_EXT_FACTOR = "{arch}.rope.scaling.yarn_ext_factor" + SCALING_YARN_ATTN_FACTOR = "{arch}.rope.scaling.yarn_attn_factor" + SCALING_YARN_BETA_FAST = "{arch}.rope.scaling.yarn_beta_fast" + SCALING_YARN_BETA_SLOW = "{arch}.rope.scaling.yarn_beta_slow" class Split: LLM_KV_SPLIT_NO = "split.no" @@ -462,6 +464,7 @@ class MODEL_ARCH(IntEnum): PANGU_EMBED = auto() MISTRAL3 = auto() MIMO2 = auto() + STEP35 = auto() LLAMA_EMBED = auto() MAINCODER = auto() KIMI_LINEAR = auto() @@ -892,6 +895,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.PANGU_EMBED: "pangu-embedded", MODEL_ARCH.MISTRAL3: "mistral3", MODEL_ARCH.MIMO2: "mimo2", + MODEL_ARCH.STEP35: "step35", MODEL_ARCH.LLAMA_EMBED: "llama-embed", MODEL_ARCH.MAINCODER: "maincoder", MODEL_ARCH.KIMI_LINEAR: "kimi-linear", @@ -3364,6 +3368,32 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_UP_EXP, MODEL_TENSOR.FFN_EXP_PROBS_B, ], + MODEL_ARCH.STEP35: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + 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_GATE, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_UP_SHEXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_EXP_PROBS_B, + ], MODEL_ARCH.LLAMA_EMBED: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, @@ -3753,12 +3783,12 @@ KEY_ATTENTION_LAYERNORM_EPS = Keys.Attention.LAYERNORM_EPS KEY_ATTENTION_LAYERNORM_RMS_EPS = Keys.Attention.LAYERNORM_RMS_EPS # RoPE -KEY_ROPE_DIMENSION_COUNT = Keys.Rope.DIMENSION_COUNT -KEY_ROPE_FREQ_BASE = Keys.Rope.FREQ_BASE -KEY_ROPE_SCALING_TYPE = Keys.Rope.SCALING_TYPE -KEY_ROPE_SCALING_FACTOR = Keys.Rope.SCALING_FACTOR -KEY_ROPE_SCALING_ORIG_CTX_LEN = Keys.Rope.SCALING_ORIG_CTX_LEN -KEY_ROPE_SCALING_FINETUNED = Keys.Rope.SCALING_FINETUNED +KEY_ROPE_DIMENSION_COUNT = Keys.Rope.DIMENSION_COUNT +KEY_ROPE_FREQ_BASE = Keys.Rope.FREQ_BASE +KEY_ROPE_SCALING_TYPE = Keys.Rope.SCALING_TYPE +KEY_ROPE_SCALING_FACTOR = Keys.Rope.SCALING_FACTOR +KEY_ROPE_SCALING_ORIG_CTX_LEN = Keys.Rope.SCALING_ORIG_CTX_LEN +KEY_ROPE_SCALING_FINETUNED = Keys.Rope.SCALING_FINETUNED # SSM KEY_SSM_CONV_KERNEL = Keys.SSM.CONV_KERNEL diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index f720aa2d54..62172b24c3 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -824,6 +824,12 @@ class GGUFWriter: def add_expert_gating_func(self, value: ExpertGatingFuncType) -> None: self.add_uint32(Keys.LLM.EXPERT_GATING_FUNC.format(arch=self.arch), value.value) + def add_swiglu_clamp_exp(self, values: Sequence[float]) -> None: + self.add_array(Keys.LLM.SWIGLU_CLAMP_EXP.format(arch=self.arch), values) + + def add_swiglu_clamp_shexp(self, values: Sequence[float]) -> None: + self.add_array(Keys.LLM.SWIGLU_CLAMP_SHEXP.format(arch=self.arch), values) + def add_expert_group_scale(self, value: float) -> None: self.add_float32(Keys.LLM.EXPERT_GROUP_SCALE.format(arch=self.arch), value) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index e16c06c2a3..167ade7803 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -359,6 +359,7 @@ class TensorNameMap: MODEL_TENSOR.ATTN_GATE: ( "model.layers.{bid}.self_attn.gate_proj", # afmoe + "model.layers.{bid}.self_attn.g_proj", # step3.5 head-wise attention gate ), # Feed-forward norm @@ -423,6 +424,7 @@ class TensorNameMap: "model.layers.{bid}.mlp.router.gate", # afmoe "layers.{bid}.gate", # mistral-large "backbone.layers.{bid}.mixer.gate", # nemotron-h-moe + "model.layers.{bid}.moe.gate", # step3.5 ), MODEL_TENSOR.FFN_GATE_INP_SHEXP: ( @@ -439,6 +441,7 @@ class TensorNameMap: "backbone.layers.{bid}.mixer.gate.e_score_correction", # nemotron-h-moe "model.layers.{bid}.mlp.e_score_correction", # exaone-moe "model.layers.{bid}.block_sparse_moe.gate.e_score_correction", # kimi + "model.layers.{bid}.moe.router_bias", # step3.5 expert selection bias ), # Feed-forward up @@ -493,6 +496,7 @@ class TensorNameMap: "model.layers.{bid}.feed_forward.experts.up_proj", # llama4 "encoder.layers.{bid}.mlp.experts.mlp.w1", # nomic-bert-moe "model.layers.{bid}.block_sparse_moe.experts.up", # smallthinker + "model.layers.{bid}.moe.up_proj", # step3.5 ), MODEL_TENSOR.FFN_UP_SHEXP: ( @@ -504,6 +508,7 @@ class TensorNameMap: "layers.{bid}.shared_experts.w3", # mistral-large "backbone.layers.{bid}.mixer.shared_experts.up_proj", # nemotron-h-moe "model.layers.{bid}.block_sparse_moe.shared_experts.up_proj", # kimi + "model.layers.{bid}.share_expert.up_proj", # step3.5 ), MODEL_TENSOR.FFN_UP_CHEXP: ( @@ -543,6 +548,7 @@ class TensorNameMap: "model.layers.{bid}.block_sparse_moe.experts.w1", # phimoe (merged) "model.layers.{bid}.feed_forward.experts.gate_proj", # llama4 "model.layers.{bid}.block_sparse_moe.experts.gate", # smallthinker + "model.layers.{bid}.moe.gate_proj", # step3.5 ), MODEL_TENSOR.FFN_GATE_SHEXP: ( @@ -552,6 +558,7 @@ class TensorNameMap: "model.layers.{bid}.mlp.shared_mlp.gate_proj", # hunyuan "layers.{bid}.shared_experts.w1", # mistral-large "model.layers.{bid}.block_sparse_moe.shared_experts.gate_proj", # kimi + "model.layers.{bid}.share_expert.gate_proj", # step3.5 ), MODEL_TENSOR.FFN_GATE_CHEXP: ( @@ -606,6 +613,7 @@ class TensorNameMap: "model.layers.{bid}.feed_forward.experts.down_proj", # llama4 "encoder.layers.{bid}.mlp.experts.mlp.w2", # nomic-bert-moe "model.layers.{bid}.block_sparse_moe.experts.down", # smallthinker + "model.layers.{bid}.moe.down_proj", # step3.5 ), MODEL_TENSOR.FFN_DOWN_SHEXP: ( @@ -617,6 +625,7 @@ class TensorNameMap: "layers.{bid}.shared_experts.w2", # mistral-large "backbone.layers.{bid}.mixer.shared_experts.down_proj", # nemotron-h-moe "model.layers.{bid}.block_sparse_moe.shared_experts.down_proj", # kimi + "model.layers.{bid}.share_expert.down_proj", # step3.5 ), MODEL_TENSOR.FFN_DOWN_CHEXP: ( diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index 5238a5e934..2115fc4255 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -135,6 +135,7 @@ add_library(llama models/stablelm.cpp models/starcoder.cpp models/starcoder2.cpp + models/step35-iswa.cpp models/t5-dec.cpp models/t5-enc.cpp models/wavtokenizer-dec.cpp diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index a8bf1c9b80..bd78f1e556 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -117,7 +117,8 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_RND1, "rnd1" }, { LLM_ARCH_PANGU_EMBED, "pangu-embedded" }, { LLM_ARCH_MISTRAL3, "mistral3" }, - { LLM_ARCH_MIMO2, "mimo2" }, + { LLM_ARCH_MIMO2, "mimo2" }, + { LLM_ARCH_STEP35, "step35" }, { LLM_ARCH_LLAMA_EMBED, "llama-embed" }, { LLM_ARCH_MAINCODER, "maincoder" }, { LLM_ARCH_KIMI_LINEAR, "kimi-linear" }, @@ -162,6 +163,8 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_EXPERT_FEED_FORWARD_LENGTH, "%s.expert_feed_forward_length" }, { LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" }, { LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH, "%s.expert_chunk_feed_forward_length" }, + { LLM_KV_SWIGLU_CLAMP_EXP, "%s.swiglu_clamp_exp" }, + { LLM_KV_SWIGLU_CLAMP_SHEXP, "%s.swiglu_clamp_shexp" }, { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" }, { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" }, { LLM_KV_EXPERT_COUNT, "%s.expert_count" }, @@ -220,21 +223,21 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" }, { LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" }, - { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, - { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" }, - { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, - { LLM_KV_ROPE_FREQ_BASE_SWA, "%s.rope.freq_base_swa" }, - { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" }, - { LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" }, - { LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" }, - { LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" }, - { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" }, - { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" }, - { LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier" }, - { LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR, "%s.rope.scaling.yarn_ext_factor" }, - { LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, "%s.rope.scaling.yarn_attn_factor" }, - { LLM_KV_ROPE_SCALING_YARN_BETA_FAST, "%s.rope.scaling.yarn_beta_fast" }, - { LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, "%s.rope.scaling.yarn_beta_slow" }, + { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, + { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" }, + { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, + { LLM_KV_ROPE_FREQ_BASE_SWA, "%s.rope.freq_base_swa" }, + { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" }, + { LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" }, + { LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" }, + { LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" }, + { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" }, + { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" }, + { LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier" }, + { LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR, "%s.rope.scaling.yarn_ext_factor" }, + { LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, "%s.rope.scaling.yarn_attn_factor" }, + { LLM_KV_ROPE_SCALING_YARN_BETA_FAST, "%s.rope.scaling.yarn_beta_fast" }, + { LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, "%s.rope.scaling.yarn_beta_slow" }, { LLM_KV_SPLIT_NO, "split.no" }, { LLM_KV_SPLIT_COUNT, "split.count" }, @@ -2279,6 +2282,35 @@ static std::set llm_get_tensor_names(llm_arch arch) { LLM_TENSOR_FFN_UP_EXPS, LLM_TENSOR_FFN_EXP_PROBS_B, }; + case LLM_ARCH_STEP35: + return { + LLM_TENSOR_TOKEN_EMBD, + LLM_TENSOR_OUTPUT_NORM, + LLM_TENSOR_OUTPUT, + LLM_TENSOR_ROPE_FREQS, + LLM_TENSOR_ROPE_FACTORS_LONG, + LLM_TENSOR_ROPE_FACTORS_SHORT, + LLM_TENSOR_ATTN_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_GATE, + LLM_TENSOR_ATTN_OUT, + LLM_TENSOR_FFN_NORM, + LLM_TENSOR_FFN_GATE, + LLM_TENSOR_FFN_DOWN, + LLM_TENSOR_FFN_UP, + LLM_TENSOR_FFN_GATE_INP, + LLM_TENSOR_FFN_GATE_EXPS, + LLM_TENSOR_FFN_DOWN_EXPS, + LLM_TENSOR_FFN_UP_EXPS, + LLM_TENSOR_FFN_GATE_SHEXP, + LLM_TENSOR_FFN_UP_SHEXP, + LLM_TENSOR_FFN_DOWN_SHEXP, + LLM_TENSOR_FFN_EXP_PROBS_B, + }; case LLM_ARCH_GPTJ: case LLM_ARCH_UNKNOWN: return { diff --git a/src/llama-arch.h b/src/llama-arch.h index f092f72834..e8263369b8 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -122,6 +122,7 @@ enum llm_arch { LLM_ARCH_PANGU_EMBED, LLM_ARCH_MISTRAL3, LLM_ARCH_MIMO2, + LLM_ARCH_STEP35, LLM_ARCH_LLAMA_EMBED, LLM_ARCH_MAINCODER, LLM_ARCH_KIMI_LINEAR, @@ -166,6 +167,8 @@ enum llm_kv { LLM_KV_EXPERT_FEED_FORWARD_LENGTH, LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH, + LLM_KV_SWIGLU_CLAMP_EXP, + LLM_KV_SWIGLU_CLAMP_SHEXP, LLM_KV_USE_PARALLEL_RESIDUAL, LLM_KV_TENSOR_DATA_LAYOUT, LLM_KV_EXPERT_COUNT, diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 165cbc0a7d..bba747d37b 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -13,6 +13,8 @@ #include #include #include +#include +#include #include void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) { @@ -1014,6 +1016,26 @@ ggml_tensor * llm_graph_context::build_ffn( switch (type_op) { case LLM_FFN_SILU: if (gate && type_gate == LLM_FFN_PAR) { + // Step35: HF clamps gate (after SiLU) and up before multiplication + if (arch == LLM_ARCH_STEP35 && il >= 0) { + const float limit = hparams.swiglu_clamp_shexp[il]; + constexpr float eps = 1e-6f; + if (limit > eps) { + ggml_tensor * gate_act = ggml_silu(ctx0, cur); + cb(gate_act, "ffn_silu", il); + gate_act = ggml_clamp(ctx0, gate_act, -INFINITY, limit); + cb(gate_act, "ffn_silu_clamped", il); + + tmp = ggml_clamp(ctx0, tmp, -limit, limit); + cb(tmp, "ffn_up_clamped", il); + + cur = ggml_mul(ctx0, gate_act, tmp); + cb(cur, "ffn_swiglu_limited", il); + type_gate = LLM_FFN_SEQ; + break; + } + } + cur = ggml_swiglu_split(ctx0, cur, tmp); cb(cur, "ffn_swiglu", il); type_gate = LLM_FFN_SEQ; @@ -1316,6 +1338,25 @@ ggml_tensor * llm_graph_context::build_moe_ffn( switch (type_op) { case LLM_FFN_SILU: if (gate_exps) { + // Step35: per-layer clamp for routed experts + if (arch == LLM_ARCH_STEP35 && il >= 0) { + const float limit = hparams.swiglu_clamp_exp[il]; + constexpr float eps = 1e-6f; + if (limit > eps) { + ggml_tensor * gate_act = ggml_silu(ctx0, cur); + cb(gate_act, "ffn_moe_silu", il); + gate_act = ggml_clamp(ctx0, gate_act, -INFINITY, limit); + cb(gate_act, "ffn_moe_silu_clamped", il); + + up = ggml_clamp(ctx0, up, -limit, limit); + cb(up, "ffn_moe_up_clamped", il); + + cur = ggml_mul(ctx0, gate_act, up); + cb(cur, "ffn_moe_swiglu_limited", il); + break; + } + } + cur = ggml_swiglu_split(ctx0, cur, up); cb(cur, "ffn_moe_swiglu", il); } else { diff --git a/src/llama-hparams.h b/src/llama-hparams.h index a435043cfe..6c695bdbf6 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -206,6 +206,11 @@ struct llama_hparams { enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE; enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE; + + // Step35: optional per-layer clamps for (Swi)GLU + std::array swiglu_clamp_exp; // clamping for expert FFN + std::array swiglu_clamp_shexp; // shared expert + // this value n_pattern means that every nth layer is dense (i.e. non-SWA) // dense_first means whether the pattern is start with a dense layer // note that if n_pattern == 0, all layers are SWA diff --git a/src/llama-kv-cache-iswa.cpp b/src/llama-kv-cache-iswa.cpp index 3a34102a23..26e2cb4270 100644 --- a/src/llama-kv-cache-iswa.cpp +++ b/src/llama-kv-cache-iswa.cpp @@ -218,7 +218,9 @@ llama_memory_context_ptr llama_kv_cache_iswa::init_update(llama_context * lctx, } bool llama_kv_cache_iswa::get_can_shift() const { - return kv_base->get_size() == kv_swa->get_size(); + return kv_base->get_can_shift() && + kv_swa->get_can_shift() && + kv_base->get_size() == kv_swa->get_size(); } void llama_kv_cache_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const { diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp index c35cd6761b..cb702b2a59 100644 --- a/src/llama-kv-cache.cpp +++ b/src/llama-kv-cache.cpp @@ -974,6 +974,10 @@ void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch & } bool llama_kv_cache::get_can_shift() const { + // Step35 uses per-layer RoPE dims; K-shift assumes a single global n_rot. + if (model.arch == LLM_ARCH_STEP35) { + return false; + } return true; } diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 765e4de2e4..674d06c891 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -130,6 +130,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_100B_A6B: return "100B.A6B"; case LLM_TYPE_102B_A12B: return "102B.A12B"; case LLM_TYPE_106B_A12B: return "106B.A12B"; + case LLM_TYPE_196B_A11B: return "196B.A11B"; case LLM_TYPE_230B_A10B: return "230B.A10B"; case LLM_TYPE_235B_A22B: return "235B.A22B"; case LLM_TYPE_300B_A47B: return "300B.A47B"; @@ -560,6 +561,8 @@ void llama_model::load_hparams(llama_model_loader & ml) { std::fill(hparams.xielu_alpha_p.begin(), hparams.xielu_alpha_p.end(), 0.0f); std::fill(hparams.xielu_beta.begin(), hparams.xielu_beta.end(), 0.0f); std::fill(hparams.xielu_eps.begin(), hparams.xielu_eps.end(), 0.0f); + std::fill(hparams.swiglu_clamp_exp.begin(), hparams.swiglu_clamp_exp.end(), 0.0f); + std::fill(hparams.swiglu_clamp_shexp.begin(), hparams.swiglu_clamp_shexp.end(), 0.0f); ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff_arr, hparams.n_layer, false); ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer, false); @@ -2482,6 +2485,35 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_STEP35: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + hparams.swa_type = LLAMA_SWA_TYPE_STANDARD; + + // MoE + SWA parameters + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false); + ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false); + + // Step35 uses sigmoid gating by default (if not set in GGUF) + if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) { + hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID; + } + + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); + ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa); + ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer); + ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_EXP, hparams.swiglu_clamp_exp, hparams.n_layer, false); + ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_SHEXP, hparams.swiglu_clamp_shexp, hparams.n_layer, false); + + switch (hparams.n_layer) { + case 45: type = LLM_TYPE_196B_A11B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; default: throw std::runtime_error("unsupported model architecture"); } @@ -7107,6 +7139,72 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED); } } break; + case LLM_ARCH_STEP35: + { + 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}, 0); + + // STEP35 supports per-layer partial RoPE dims; rope factors are stored as a single shared tensor + // ("rope_freqs.weight") and ggml uses only the first (n_rot_l/2) entries per layer. + uint32_t n_rot_max = 0; + for (int i = 0; i < n_layer; ++i) { + n_rot_max = std::max(n_rot_max, hparams.n_rot); + } + if (n_rot_max == 0) { + n_rot_max = n_rot; + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + const uint32_t n_head_l = hparams.n_head(i); + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i); + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED); + + // optional rope factors (llama3) / longrope tensors + if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) { + layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + } else { + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + } + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head_l}, 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_v * n_head_l, n_embd}, 0); + + // head-wise attention gate (Step35 self_attn.g_proj) + layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head_l}, TENSOR_NOT_REQUIRED); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + // dense MLP (leading dense blocks) + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, TENSOR_NOT_REQUIRED); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED); + + // MoE routed experts + selection bias (router_bias) + const int64_t n_ff_exp = hparams.n_ff_exp; + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED); + + // shared expert MLP + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, TENSOR_NOT_REQUIRED); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, TENSOR_NOT_REQUIRED); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED); + } + } break; case LLM_ARCH_MAINCODER: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -8257,6 +8355,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { { llm = std::make_unique(*this, params); } break; + case LLM_ARCH_STEP35: + { + llm = std::make_unique(*this, params); + } break; default: GGML_ABORT("fatal error"); } @@ -8502,6 +8604,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_AFMOE: case LLM_ARCH_QWEN3NEXT: case LLM_ARCH_MIMO2: + case LLM_ARCH_STEP35: return LLAMA_ROPE_TYPE_NEOX; case LLM_ARCH_QWEN2VL: diff --git a/src/llama-model.h b/src/llama-model.h index 5b408bcea2..7b580043b3 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -123,6 +123,7 @@ enum llm_type { LLM_TYPE_100B_A6B, LLM_TYPE_102B_A12B, // Solar-Open LLM_TYPE_106B_A12B, // GLM-4.5-Air + LLM_TYPE_196B_A11B, // Step3.5-Flash LLM_TYPE_230B_A10B, // Minimax M2 LLM_TYPE_235B_A22B, LLM_TYPE_300B_A47B, // Ernie MoE big diff --git a/src/models/models.h b/src/models/models.h index 71c1fe8108..cfcbb9aaa5 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -583,6 +583,10 @@ struct llm_build_starcoder : public llm_graph_context { llm_build_starcoder(const llama_model & model, const llm_graph_params & params); }; +struct llm_build_step35_iswa : public llm_graph_context { + llm_build_step35_iswa(const llama_model & model, const llm_graph_params & params); +}; + struct llm_build_t5_dec : public llm_graph_context { llm_build_t5_dec(const llama_model & model, const llm_graph_params & params); }; diff --git a/src/models/step35-iswa.cpp b/src/models/step35-iswa.cpp new file mode 100644 index 0000000000..f8737815a6 --- /dev/null +++ b/src/models/step35-iswa.cpp @@ -0,0 +1,168 @@ +#include "models.h" + +llm_build_step35_iswa::llm_build_step35_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + ggml_tensor * inp_pos = build_inp_pos(); + auto * inp_attn = build_attn_inp_kv_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + const uint32_t n_head_l = hparams.n_head(il); + const uint32_t n_head_kv_l = hparams.n_head_kv(il); + + const float freq_base_l = model.get_rope_freq_base(cparams, il); + const float freq_scale_l = model.get_rope_freq_scale(cparams, il); + + cur = inpL; + + // dump pre-attn RMSNorm input to pinpoint layer boundary issues + cb(cur, "attn_norm_in", il); + + // self-attention + { + cur = build_norm(cur, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens); + + // Q/K per-head RMSNorm (Step35 q_norm / k_norm) + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + } + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + } + + // RoPE (partial rotary factors per layer) + const bool is_swa = hparams.is_swa(il); + ggml_tensor * rope_factors = is_swa ? nullptr : model.get_rope_factors(cparams, il); + const int64_t n_rot_l = is_swa ? hparams.n_rot : (hparams.n_rot / 2); + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow + ); + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur_pos", il); + cb(Kcur, "Kcur_pos", il); + + const float kq_scale = 1.0f / sqrtf(float(n_embd_head_k)); + ggml_tensor * attn_out = build_attn(inp_attn, + nullptr, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(attn_out, "attn_out", il); + // head-wise attention gate: sigmoid(g_proj(x)) in torch + if (model.layers[il].wqkv_gate) { + ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, cur); // [n_head_l, n_tokens] + cb(gate, "attn_gate", il); + + gate = ggml_sigmoid(ctx0, gate); + cb(gate, "attn_gate_sigmoid", il); + + // reshape + broadcast to [n_embd_head_v, n_head_l, n_tokens] + ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, attn_out, n_embd_head_v, n_head_l, n_tokens); + ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate, 1, n_head_l, n_tokens); + cb(gate_3d, "attn_gate_3d", il); + + attn_3d = ggml_mul(ctx0, attn_3d, gate_3d); + cb(attn_3d, "attn_gated_3d", il); + + attn_out = ggml_reshape_2d(ctx0, attn_3d, n_embd_head_v * n_head_l, n_tokens); + cb(attn_out, "attn_gated", il); + } + + // output projection + cur = build_lora_mm(model.layers[il].wo, attn_out); + cb(cur, "attn_proj", 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); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward + if (model.layers[il].ffn_gate_inp == nullptr) { + // dense MLP + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, nullptr, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, nullptr, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, nullptr, + nullptr, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE routed experts + const bool norm_w = hparams.expert_weights_norm; + const float w_scale = hparams.expert_weights_scale; + const bool scale_w = w_scale != 0.0f; + 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, + model.layers[il].ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, + norm_w, scale_w, w_scale, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il); + cb(moe_out, "ffn_moe_out", il); + + // shared expert MLP (always added on MoE layers in Step35) + ggml_tensor * sh_out = build_ffn(cur, + model.layers[il].ffn_up_shexp, nullptr, nullptr, + model.layers[il].ffn_gate_shexp, nullptr, nullptr, + model.layers[il].ffn_down_shexp, nullptr, nullptr, + nullptr, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(sh_out, "ffn_shared_out", il); + + cur = ggml_add(ctx0, moe_out, sh_out); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1); + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +}