diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 2419126ec4ea2..acd2e89e5ae1d 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -4890,6 +4890,9 @@ def __init__(self, dir_model: Path, *args, **kwargs): with open(dir_model / "config.json", "r", encoding="utf-8") as f: hparams = json.load(f) super().__init__(dir_model, *args, hparams=hparams, **kwargs) + self.d_model = self.find_hparam(["hidden_size", "d_model", "dim"]) + self.d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * self.d_model + self.n_group = self.find_hparam(["n_groups"], optional=True) or 1 def set_vocab(self): vocab_size = self.hparams["vocab_size"] @@ -4912,12 +4915,9 @@ def set_vocab(self): self._set_vocab_builtin("gpt-neox", vocab_size) def set_gguf_parameters(self): - d_model = self.find_hparam(["hidden_size", "d_model", "dim"]) - d_conv = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4 - d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * d_model - d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 128 - head_dim = self.find_hparam(["mamba_d_head", "head_dim"], optional=True) or 64 - n_group = self.find_hparam(["n_groups"], optional=True) or 1 + d_conv = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4 + d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 128 + head_dim = self.find_hparam(["mamba_d_head", "head_dim"], optional=True) or 64 rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5 @@ -4925,19 +4925,19 @@ def set_gguf_parameters(self): # TODO: does this really matter? # skip the assertion for FalconH1 Model if self.model_arch != gguf.MODEL_ARCH.FALCON_H1: - assert d_inner == 2 * d_model - assert d_inner % head_dim == 0 + assert self.d_inner == 2 * self.d_model + assert self.d_inner % head_dim == 0 self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default - self.gguf_writer.add_embedding_length(d_model) + self.gguf_writer.add_embedding_length(self.d_model) self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading self.gguf_writer.add_block_count(self.block_count) self.gguf_writer.add_ssm_conv_kernel(d_conv) - self.gguf_writer.add_ssm_inner_size(d_inner) + self.gguf_writer.add_ssm_inner_size(self.d_inner) self.gguf_writer.add_ssm_state_size(d_state) - self.gguf_writer.add_ssm_time_step_rank(d_inner // head_dim) - self.gguf_writer.add_ssm_group_count(n_group) + self.gguf_writer.add_ssm_time_step_rank(self.d_inner // head_dim) + self.gguf_writer.add_ssm_group_count(self.n_group) self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps) self.gguf_writer.add_file_type(self.ftype) @@ -4962,10 +4962,7 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter # (D is also unsqueezed, but for more straightforward broadcast internally) data_torch = data_torch.reshape((*data_torch.shape, 1)) elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_NORM, bid): - d_model = self.find_hparam(["hidden_size", "d_model", "dim"]) - d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * d_model - n_group = self.hparams.get("n_groups", 1) - data_torch = data_torch.reshape((n_group, d_inner // n_group)) + data_torch = data_torch.reshape((self.n_group, self.d_inner // self.n_group)) if name.endswith(".A_log"): logger.debug("A_log --> A ==> " + new_name) @@ -6452,18 +6449,152 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up), ] + has_experts = bool(self.hparams.get('num_local_experts')) + if name.endswith("shared_mlp.input_linear.weight"): ffn_dim = self.hparams["shared_intermediate_size"] assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size" gate, up = data_torch.split(ffn_dim, dim=-2) + if has_experts: + return [ + (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate), + (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up), + ] + return [ + (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), gate), + (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), up), + ] + + if not has_experts and name.endswith("shared_mlp.output_linear.weight"): return [ - (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate), - (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up), + (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid), data_torch) ] return super().modify_tensors(data_torch, name, bid) +@ModelBase.register("GraniteMoeHybridForCausalLM", "BambaForCausalLM") +class GraniteHybridModel(Mamba2Model, GraniteMoeModel): + """GraniteHybrid is a hybrid SSM + Attention model that uses Mamba2 SSM + layers and optionally uses MoE w/ a shared expert""" + model_arch = gguf.MODEL_ARCH.GRANITE_HYBRID + undo_permute = True + + def __init__(self, *args, **kwargs): + + # Hybrid mamba models use a prefix for the mamba-specific params. + # TODO: Extend this if the prefix(es) need to be configurable + self.hparam_prefixes = ["mamba"] + + super().__init__(*args, **kwargs) + + # Use Granite conversion for attention + self._transformer_model_class: type[TextModel] = GraniteModel + + # Lists of which layers use ssm vs attention + self._attn_layers = self.get_attn_layres() + self._ssm_layers = [ + i for i in range(self.block_count) + if i not in self._attn_layers + ] + + # n_group and d_inner are used during reshape_tensors for mamaba2 + self.d_model = self.find_hparam(["hidden_size", "d_model"]) + self.n_group = self.find_hparam(["n_groups"]) + self.d_inner = self.find_hparam(["expand"]) * self.d_model + + def get_attn_layres(self): + # Explicit list of layer type names + if layer_types := self.hparams.get("layer_types"): + return [ + i for i, typ in enumerate(layer_types) + if typ == "attention" + ] + + # Layer types indicated by index or period + attn_layers = self.hparams.get("attn_layer_indices", []) + if not attn_layers: + attn_period = self.hparams.get("attn_layer_period") + assert attn_period, "Didn't find attn_layer_indices or attn_layer_period" + attn_offset = self.hparams.get("attn_layer_offset") + assert attn_offset is not None, "No attention layer offset set with attn_layer_period" + attn_layers = [ + i for i in range(self.block_count) + if i % attn_period == attn_offset + ] + return attn_layers + + def find_hparam(self, keys: Iterable[str], *args, **kwargs) -> Any: + prefixed = [] + for pfx in self.hparam_prefixes: + prefixed.extend( + "_".join([pfx, k]) + for k in keys + ) + keys = list(keys) + prefixed + return super().find_hparam(keys, *args, **kwargs) + + def modify_tensors( + self, data_torch: Tensor, name: str, bid: int | None + ) -> Iterable[tuple[str, Tensor]]: + if ( + name.endswith("block_sparse_moe.input_linear.weight") + or "shared_mlp" in name + ): + return GraniteMoeModel.modify_tensors(self, data_torch, name, bid) + + # Determine whether this is a mamaba layer or an attention layer + if bid in self._ssm_layers: + return super().modify_tensors(data_torch, name, bid) + elif bid in self._attn_layers: + return self._transformer_model_class.modify_tensors(self, data_torch, name, bid) + return [(self.map_tensor_name(name), data_torch)] + + def set_gguf_parameters(self): + GraniteMoeModel.set_gguf_parameters(self) + + ## General Params ## + self.gguf_writer.add_embedding_length(self.d_model) + self.gguf_writer.add_block_count(self.block_count) + self.gguf_writer.add_context_length(self.hparams.get("max_position_embeddings", 0)) + self.gguf_writer.add_vocab_size(self.hparams["vocab_size"]) + self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) + + ## Mamba mixer params ## + self.gguf_writer.add_ssm_conv_kernel(self.find_hparam(["conv_kernel", "d_conv"])) + self.gguf_writer.add_ssm_state_size(self.find_hparam(["state_size", "d_state"])) + self.gguf_writer.add_ssm_group_count(self.n_group) + self.gguf_writer.add_ssm_inner_size(self.d_inner) + # NOTE: The mamba_dt_rank is _not_ the right field for how this is used + # in llama.cpp + self.gguf_writer.add_ssm_time_step_rank(self.find_hparam(["n_heads"])) + + ## Attention params ## + self.gguf_writer.add_attn_layer_indices(self._attn_layers) + if rope_dim := self.hparams.get("attn_rotary_emb"): + self.gguf_writer.add_rope_dimension_count(rope_dim) + self.gguf_writer.add_head_count(self.hparams["num_attention_heads"]) + self.gguf_writer.add_head_count_kv(self.find_hparam(["num_key_value_heads", "n_head_kv"])) + + ## Feed Forward Params ## + self.gguf_writer.add_layer_norm_rms_eps( + self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5 + ) + + ## If Bamba, use rope, otherwise don't + use_rope = "BambaForCausalLM" in self.hparams["architectures"] + self.gguf_writer.add_rope_scaling_finetuned(use_rope) + + ## Validation ## + d_head = self.find_hparam(["d_head"], optional=True) or 64 + assert self.hparams.get("hidden_act") in [None, "silu"], "Only SILU activation supported" + assert self.d_inner % d_head == 0, f"SSM inner size {self.d_inner} not a multiple of head dim {d_head}" + + def set_vocab(self): + self.hparams["pad_vocab_size_multiple"] = 8 + super().set_vocab() + + @ModelBase.register("BailingMoeForCausalLM") class BailingMoeModel(TextModel): model_arch = gguf.MODEL_ARCH.BAILINGMOE diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index fbe3f53273a35..a672e574ccbd7 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -173,6 +173,9 @@ class SSM: GROUP_COUNT = "{arch}.ssm.group_count" DT_B_C_RMS = "{arch}.ssm.dt_b_c_rms" + class HybridAttention: + ATTN_LAYER_INDICES = "{arch}.attention.layer_indices" + class WKV: HEAD_SIZE = "{arch}.wkv.head_size" @@ -352,6 +355,7 @@ class MODEL_ARCH(IntEnum): EXAONE = auto() GRANITE = auto() GRANITE_MOE = auto() + GRANITE_HYBRID = auto() CHAMELEON = auto() WAVTOKENIZER_DEC = auto() PLM = auto() @@ -661,6 +665,7 @@ class MODEL_TENSOR(IntEnum): MODEL_ARCH.EXAONE: "exaone", MODEL_ARCH.GRANITE: "granite", MODEL_ARCH.GRANITE_MOE: "granitemoe", + MODEL_ARCH.GRANITE_HYBRID: "granitehybrid", MODEL_ARCH.CHAMELEON: "chameleon", MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec", MODEL_ARCH.PLM: "plm", @@ -2143,6 +2148,36 @@ class MODEL_TENSOR(IntEnum): MODEL_TENSOR.FFN_UP_SHEXP, MODEL_TENSOR.FFN_DOWN_SHEXP, ], + MODEL_ARCH.GRANITE_HYBRID: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.SSM_IN, + MODEL_TENSOR.SSM_CONV1D, + MODEL_TENSOR.SSM_DT, + MODEL_TENSOR.SSM_A, + MODEL_TENSOR.SSM_D, + MODEL_TENSOR.SSM_NORM, + MODEL_TENSOR.SSM_OUT, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + # MoE + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + # Dense + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], MODEL_ARCH.CHAMELEON: [ 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 a7ecf3d31209f..6aca08b5fc378 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -270,7 +270,14 @@ def write_ti_data_to_file(self) -> None: self.state = WriterState.TI_DATA def add_key_value(self, key: str, val: Any, vtype: GGUFValueType, sub_type: GGUFValueType | None = None) -> None: - if any(key in kv_data for kv_data in self.kv_data): + # Warn about duplicate keys if they differ by value or type + if any( + ( + key in kv_data + and (kv_data[key].value != val or kv_data[key].type != vtype) + ) + for kv_data in self.kv_data + ): logger.warning(f'Duplicated key name {key!r}, overwriting it with new value {val!r} of type {vtype.name}') self.kv_data[0][key] = GGUFValue(value=val, type=vtype, sub_type=sub_type) @@ -867,6 +874,9 @@ def add_ssm_group_count(self, value: int) -> None: def add_ssm_dt_b_c_rms(self, value: bool) -> None: self.add_bool(Keys.SSM.DT_B_C_RMS.format(arch=self.arch), value) + def add_attn_layer_indices(self, values: list[int]) -> None: + self.add_array(Keys.HybridAttention.ATTN_LAYER_INDICES.format(arch=self.arch), values) + def add_tokenizer_model(self, model: str) -> None: self.add_string(Keys.Tokenizer.MODEL, model) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 215eb297ebcc1..7a4f275ceec28 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -13,7 +13,7 @@ class TensorNameMap: "transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais exaone "transformer.word_embeddings", # falcon "word_embeddings", # bloom - "model.embed_tokens", # llama-hf nemotron olmoe olmo2 rwkv6qwen2 glm4-0414 + "model.embed_tokens", # llama-hf nemotron olmoe olmo2 rwkv6qwen2 glm4-0414 granite-hybrid "tok_embeddings", # llama-pth "embeddings.word_embeddings", # bert nomic-bert "language_model.embedding.word_embeddings", # persimmon @@ -118,7 +118,7 @@ class TensorNameMap: "transformer.h.{bid}.input_layernorm", # falcon7b "h.{bid}.input_layernorm", # bloom "transformer.h.{bid}.ln_mlp", # falcon40b - "model.layers.{bid}.input_layernorm", # llama-hf nemotron olmoe phimoe + "model.layers.{bid}.input_layernorm", # llama-hf nemotron olmoe phimoe granite-hybrid "layers.{bid}.attention_norm", # llama-pth "language_model.encoder.layers.{bid}.input_layernorm", # persimmon "model.layers.{bid}.ln1", # yi @@ -279,7 +279,7 @@ class TensorNameMap: "transformer.decoder_layer.{bid}.rms_norm_2", # Grok "encoder.layers.{bid}.post_attention_layernorm", # chatglm "transformer.layers.{bid}.ffn_norm", # openelm - "model.layers.{bid}.pre_ff_layernorm", # jamba + "model.layers.{bid}.pre_ff_layernorm", # jamba granite-hybrid "model.layers.{bid}.pre_moe_layernorm", # mini-jamba "model.layers.{bid}.post_attention_layernorm", # llama4 "transformer_encoder.{bid}.ffn_norm", # neobert @@ -349,7 +349,7 @@ class TensorNameMap: "model.layers.{bid}.residual_mlp.w3", # arctic "encoder.layers.{bid}.mlp.dense_h_to_4h", # chatglm "transformer.h.{bid}.mlp.c_fc_1", # exaone - "model.layers.{bid}.feed_forward.up_proj", # llama4 jamba + "model.layers.{bid}.feed_forward.up_proj", # llama4 jamba granite-hybrid "transformer_encoder.{bid}.ffn.w12", # neobert ), @@ -389,7 +389,7 @@ class TensorNameMap: "transformer.h.{bid}.mlp.linear_1", # refact "model.layers.{bid}.residual_mlp.w1", # arctic "transformer.h.{bid}.mlp.c_fc_0", # exaone - "model.layers.{bid}.feed_forward.gate_proj", # llama4 jamba + "model.layers.{bid}.feed_forward.gate_proj", # llama4 jamba granite-hybrid ), MODEL_TENSOR.FFN_GATE_EXP: ( @@ -435,7 +435,7 @@ class TensorNameMap: "encoder.layer.{bid}.mlp.down_layer", # jina-bert-v2 "encoder.layers.{bid}.mlp.dense_4h_to_h", # chatglm "model.layers.h.{bid}.mlp.c_proj", # exaone - "model.layers.{bid}.feed_forward.down_proj", # llama4 jamba + "model.layers.{bid}.feed_forward.down_proj", # llama4 jamba granite-hybrid "transformer_encoder.{bid}.ffn.w3", # neobert ), @@ -558,13 +558,13 @@ class TensorNameMap: MODEL_TENSOR.SSM_IN: ( "model.layers.{bid}.in_proj", # mamba-hf "backbone.layers.{bid}.mixer.in_proj", # mamba - "model.layers.{bid}.mamba.in_proj", # jamba falcon-h1 + "model.layers.{bid}.mamba.in_proj", # jamba falcon-h1 granite-hybrid ), MODEL_TENSOR.SSM_CONV1D: ( "model.layers.{bid}.conv1d", # mamba-hf "backbone.layers.{bid}.mixer.conv1d", # mamba - "model.layers.{bid}.mamba.conv1d", # jamba falcon-h1 + "model.layers.{bid}.mamba.conv1d", # jamba falcon-h1 granite-hybrid ), MODEL_TENSOR.SSM_X: ( @@ -576,7 +576,7 @@ class TensorNameMap: MODEL_TENSOR.SSM_DT: ( "model.layers.{bid}.dt_proj", # mamba-hf "backbone.layers.{bid}.mixer.dt_proj", # mamba - "model.layers.{bid}.mamba.dt_proj", # jamba falcon-h1 + "model.layers.{bid}.mamba.dt_proj", # jamba falcon-h1 granite-hybrid ), MODEL_TENSOR.SSM_DT_NORM: ( @@ -586,7 +586,7 @@ class TensorNameMap: MODEL_TENSOR.SSM_A: ( "model.layers.{bid}.A_log", # mamba-hf "backbone.layers.{bid}.mixer.A_log", # mamba - "model.layers.{bid}.mamba.A_log", # jamba falcon-h1 + "model.layers.{bid}.mamba.A_log", # jamba falcon-h1 granite-hybrid ), MODEL_TENSOR.SSM_B_NORM: ( @@ -602,18 +602,18 @@ class TensorNameMap: MODEL_TENSOR.SSM_D: ( "model.layers.{bid}.D", # mamba-hf "backbone.layers.{bid}.mixer.D", # mamba - "model.layers.{bid}.mamba.D", # jamba falcon-h1 + "model.layers.{bid}.mamba.D", # jamba falcon-h1 granite-hybrid ), MODEL_TENSOR.SSM_NORM: ( - "model.layers.{bid}.mamba.norm", # falcon-h1 + "model.layers.{bid}.mamba.norm", # falcon-h1 granite-hybrid "backbone.layers.{bid}.mixer.norm", # mamba2 ), MODEL_TENSOR.SSM_OUT: ( "model.layers.{bid}.out_proj", # mamba-hf "backbone.layers.{bid}.mixer.out_proj", # mamba - "model.layers.{bid}.mamba.out_proj", # jamba falcon-h1 + "model.layers.{bid}.mamba.out_proj", # jamba falcon-h1 granite-hybrid ), MODEL_TENSOR.TIME_MIX_W0: ( diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 1955c03eb3d1c..e60c408601611 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -73,6 +73,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_ARWKV7, "arwkv7" }, { LLM_ARCH_GRANITE, "granite" }, { LLM_ARCH_GRANITE_MOE, "granitemoe" }, + { LLM_ARCH_GRANITE_HYBRID, "granitehybrid" }, { LLM_ARCH_CHAMELEON, "chameleon" }, { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" }, { LLM_ARCH_PLM, "plm" }, @@ -1641,6 +1642,43 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, }, }, + { + LLM_ARCH_GRANITE_HYBRID, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + // mamba(2) ssm layers + { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" }, + { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" }, + { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" }, + { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" }, + { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" }, + { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" }, + { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, + // attention layers + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + // dense FFN + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + // moe FFN + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + // shared expert + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + }, + }, { LLM_ARCH_CHAMELEON, { @@ -2027,10 +2065,10 @@ bool llm_arch_is_recurrent(const llm_arch & arch) { } bool llm_arch_is_hybrid(const llm_arch & arch) { - // List all mamba-attention hybrid models here switch (arch) { case LLM_ARCH_JAMBA: case LLM_ARCH_FALCON_H1: + case LLM_ARCH_GRANITE_HYBRID: return true; default: return false; diff --git a/src/llama-arch.h b/src/llama-arch.h index 3381b8dc4a4b7..a82af4032bff2 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -77,6 +77,7 @@ enum llm_arch { LLM_ARCH_ARWKV7, LLM_ARCH_GRANITE, LLM_ARCH_GRANITE_MOE, + LLM_ARCH_GRANITE_HYBRID, LLM_ARCH_CHAMELEON, LLM_ARCH_WAVTOKENIZER_DEC, LLM_ARCH_PLM, diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp index bd9e6da8832b7..0bd1e5d006950 100644 --- a/src/llama-model-loader.cpp +++ b/src/llama-model-loader.cpp @@ -464,6 +464,7 @@ namespace GGUFMeta { // TODO: this is not very clever - figure out something better template bool llama_model_loader::get_key_or_arr>(enum llm_kv kid, std::array & result, uint32_t n, bool required); template bool llama_model_loader::get_key_or_arr>(enum llm_kv kid, std::array & result, uint32_t n, bool required); + template bool llama_model_loader::get_arr(enum llm_kv kid, std::vector & result, bool required); llama_model_loader::llama_model_loader( const std::string & fname, diff --git a/src/llama-model.cpp b/src/llama-model.cpp index ca094e47b6cb5..e27263faf700b 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -1504,6 +1504,11 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale); ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale); + // Granite uses rope_finetuned as a switch for rope, so default to true + bool rope_finetuned = true; + ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false); + hparams.rope_finetuned = rope_finetuned; + switch (hparams.n_layer) { case 32: type = LLM_TYPE_3B; break; case 40: type = LLM_TYPE_3B; break; @@ -1511,6 +1516,57 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } + // For Granite MoE Shared + ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false); + } break; + case LLM_ARCH_GRANITE_HYBRID: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale, /* required */ false); + ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale, /* required */ false); + ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale, /* required */ false); + ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale, /* required */ false); + + 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); + + // Granite uses rope_finetuned as a switch for rope, so default to true + bool rope_finetuned = true; + ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false); + hparams.rope_finetuned = rope_finetuned; + + // Zero-out n_head_arr and n_head_kv_arr since SSM layers don't + // have attention heads. We'll set them correctly below once we + // know which layers are attention layers + // NOTE: It's important that this happens after n_embd_head_[kv] + // are set above! + const auto n_head_attn = hparams.n_head(); + const auto n_head_kv_attn = hparams.n_head_kv(); + std::fill(hparams.n_head_arr.begin(), hparams.n_head_arr.end(), 0); + std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0); + + // Attention params + std::fill(hparams.recurrent_layer_arr.begin(), hparams.recurrent_layer_arr.end(), true); + std::vector attn_layer_indices; + ml.get_arr(LLM_KV_ATTENTION_LAYER_INDICES, attn_layer_indices); + for (const auto attn_idx : attn_layer_indices) { + GGML_ASSERT(attn_idx < hparams.n_layer); + hparams.recurrent_layer_arr[attn_idx] = false; + // Correctly set n_head and n_head_kv for attention layers + hparams.n_head_arr[attn_idx] = n_head_attn; + hparams.n_head_kv_arr[attn_idx] = n_head_kv_attn; + } + + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + // TODO: Add llm type label (not sure this is useful) + default: type = LLM_TYPE_UNKNOWN; + } + // For Granite MoE Shared ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false); } break; @@ -3362,6 +3418,100 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } } } break; + case LLM_ARCH_GRANITE_HYBRID: + { + // mamba2 Mixer SSM params + // NOTE: int64_t for tensor dimensions + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t n_ssm_head = hparams.ssm_dt_rank; + const int64_t n_group = hparams.ssm_n_group; + const int64_t d_in_proj = 2*d_inner + 2*n_group*d_state + n_ssm_head; + + // only an expansion factor of 2 is supported for now + GGML_ASSERT(2 * n_embd == d_inner); + + // embeddings + 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, duplicated to allow offloading + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + // norm + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + if (hparams.is_recurrent(i)) { + // ssm layers + layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0); + layer.ssm_in_b = create_tensor(tn(LLM_TENSOR_SSM_IN, "bias", i), {n_embd, d_in_proj}, TENSOR_NOT_REQUIRED); + + layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0); + layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, TENSOR_NOT_REQUIRED); + + layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, 0); + + // no "weight" suffix for these + layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, 0); + layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, 0); + + layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0); + + // out_proj + layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0); + } else { + // attention layers (with optional bias) + const int64_t n_head_i = hparams.n_head(i); + const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i); + const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i); + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head_i}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa_i}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa_i}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0); + layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); + layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_k_gqa_i}, TENSOR_NOT_REQUIRED); + layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_v_gqa_i}, TENSOR_NOT_REQUIRED); + layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); + } + + // feed forward (w/ optional biases) + if (n_expert > 0) { + // MoE FFN + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 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, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); + + // For Granite MoE Shared + if (hparams.n_ff_shexp > 0) { + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0); + } + } else { + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 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); + layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED); + layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); + layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED); + } + } + } break; case LLM_ARCH_XVERSE: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -5079,7 +5229,8 @@ void llama_model::print_info() const { if (arch == LLM_ARCH_MINICPM || arch == LLM_ARCH_GRANITE || - arch == LLM_ARCH_GRANITE_MOE) { + arch == LLM_ARCH_GRANITE_MOE || + arch == LLM_ARCH_GRANITE_HYBRID) { LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale); LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale); LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale); @@ -13795,13 +13946,11 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base { } }; - struct llm_build_granite : public llm_graph_context { llm_build_granite( const llama_model & model, const llm_graph_params & params, - ggml_cgraph * gf, - const bool use_rope = true) + ggml_cgraph * gf) : llm_graph_context(params) { const int64_t n_embd_head = hparams.n_embd_head_v; @@ -13816,14 +13965,12 @@ struct llm_build_granite : public llm_graph_context { // inp_pos - built only if rope enabled ggml_tensor * inp_pos = nullptr; - if (use_rope) { + if (hparams.rope_finetuned) { inp_pos = build_inp_pos(); } auto * inp_attn = build_attn_inp_kv_unified(); - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - ggml_tensor * inp_out_ids = build_inp_out_ids(); for (int il = 0; il < n_layer; ++il) { @@ -13836,128 +13983,237 @@ struct llm_build_granite : public llm_graph_context { cb(cur, "attn_norm", il); // self-attention - { - // compute Q and K and (optionally) RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } + cur = build_attention_layer( + gf, cur, inp_pos, inp_attn, + model, n_embd_head, il); - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", 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 * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } + // ffn + cur = build_layer_ffn(cur, inpSA, model, il); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + // input for next layer + inpL = cur; + } - if (use_rope) { - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); + cur = inpL; - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - } + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); + cb(cur, "result_norm", -1); + res->t_embd = cur; - cur = build_attn(inp_attn, gf, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); + // lm_head + cur = build_lora_mm(model.output, cur); + + // For Granite architectures - scale logits + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } + + ggml_tensor * build_attention_layer( + ggml_cgraph * gf, + ggml_tensor * cur, + ggml_tensor * inp_pos, + llm_graph_input_attn_kv_unified * inp_attn, + const llama_model & model, + const int64_t n_embd_head, + const int il) { + + // compute Q and K and (optionally) RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + + const bool use_rope = hparams.rope_finetuned; + if (use_rope) { + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, 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); + + 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_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); cb(cur, "attn_out", il); - } + return cur; + } - 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 * build_layer_ffn( + ggml_tensor * cur, + ggml_tensor * inpSA, + const llama_model & model, + const int il) { - // For Granite architectures - scale residual + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); - // feed-forward network (non-MoE) - if (model.layers[il].ffn_gate_inp == nullptr) { + // feed-forward network (non-MoE) + if (model.layers[il].ffn_gate_inp == nullptr) { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); - } else { - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); + } else { + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); - 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); + 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); - // For Granite MoE Shared - if (hparams.n_ff_shexp > 0) { - 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); + // For Granite MoE Shared + if (hparams.n_ff_shexp > 0) { + 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); - cur = ggml_add(ctx0, moe_out, ffn_shexp); - cb(cur, "ffn_out", il); - } else { - cur = moe_out; - } + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } else { + cur = moe_out; } + } - // For Granite architectures - scale residual + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); - cur = build_cvec(cur, il); - cb(cur, "l_out", il); + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + return cur; + } +}; + +struct llm_build_granite_hybrid : public llm_graph_context_mamba { + + llm_build_granite_hybrid( + const llama_model & model, + const llm_graph_params & params, + ggml_cgraph * gf) : + llm_graph_context_mamba(params) { + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp = build_inp_mem_hybrid(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + // Positional embeddings populated if rope enabled + ggml_tensor * inp_pos = nullptr; + if (hparams.rope_finetuned) { + inp_pos = build_inp_pos(); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + if (hparams.is_recurrent(il)) { + // ssm layer // + cur = build_mamba2_layer(inp->get_recr(), gf, cur, model, ubatch, il); + } else { + // attention layer // + cur = build_attention_layer( + gf, cur, inp_pos, inp->get_attn(), model, + n_embd_head, 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); + } + + // ffn + cur = build_layer_ffn(cur, inpSA, model, il); // input for next layer inpL = cur; @@ -13976,12 +14232,156 @@ struct llm_build_granite : public llm_graph_context { cur = build_lora_mm(model.output, cur); // For Granite architectures - scale logits - cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); + if (hparams.f_logit_scale) { + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); + } cb(cur, "result_output", -1); res->t_logits = cur; ggml_build_forward_expand(gf, cur); } + + ggml_tensor * build_attention_layer( + ggml_cgraph * gf, + ggml_tensor * cur, + ggml_tensor * inp_pos, + llm_graph_input_attn_kv_unified * inp_attn, + const llama_model & model, + const int64_t n_embd_head, + const int il) { + + // compute Q and K and (optionally) RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + + const bool use_rope = hparams.rope_finetuned; + if (use_rope) { + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, 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); + + 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_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + return cur; + } + + ggml_tensor * build_layer_ffn( + ggml_tensor * cur, + ggml_tensor * inpSA, + const llama_model & model, + const int il) { + + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network (non-MoE) + if (model.layers[il].ffn_gate_inp == nullptr) { + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + } else { + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + 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); + + // For Granite MoE Shared + if (hparams.n_ff_shexp > 0) { + 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); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } else { + cur = moe_out; + } + } + + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + return cur; + } }; // ref: https://github.com/facebookresearch/chameleon @@ -15832,6 +16232,10 @@ llm_graph_result_ptr llama_model::build_graph( { llm = std::make_unique(*this, params, gf); } break; + case LLM_ARCH_GRANITE_HYBRID: + { + llm = std::make_unique(*this, params, gf); + } break; case LLM_ARCH_CHAMELEON: { llm = std::make_unique(*this, params, gf); @@ -16021,6 +16425,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_GLM4: case LLM_ARCH_GRANITE: case LLM_ARCH_GRANITE_MOE: + case LLM_ARCH_GRANITE_HYBRID: case LLM_ARCH_CHAMELEON: case LLM_ARCH_BAILINGMOE: case LLM_ARCH_NEO_BERT: diff --git a/src/llama-model.h b/src/llama-model.h index 453f5af62fbc7..0cafb1b12dc5c 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -261,6 +261,7 @@ struct llama_layer { // mamba bias struct ggml_tensor * ssm_conv1d_b = nullptr; struct ggml_tensor * ssm_dt_b = nullptr; + struct ggml_tensor * ssm_in_b = nullptr; // rwkv struct ggml_tensor * time_mix_w1 = nullptr;