diff --git a/conversion/__init__.py b/conversion/__init__.py index a4359f21f..fcbfb6a3a 100644 --- a/conversion/__init__.py +++ b/conversion/__init__.py @@ -20,6 +20,7 @@ TEXT_MODEL_MAP: dict[str, str] = { "NEOChatModel": "sensenova_u1", "AfmoeForCausalLM": "afmoe", "LagunaForCausalLM": "laguna", + "RicursiveForCausalLM": "ricursive", "FlintForCausalLM": "flint", "ApertusForCausalLM": "llama", "ArceeForCausalLM": "llama", diff --git a/conversion/ricursive.py b/conversion/ricursive.py new file mode 100644 index 000000000..13cb77367 --- /dev/null +++ b/conversion/ricursive.py @@ -0,0 +1,92 @@ +from __future__ import annotations + +import math + +from typing import Iterable, TYPE_CHECKING + +if TYPE_CHECKING: + from torch import Tensor + +from .base import ModelBase, TextModel, gguf, logger + + +@ModelBase.register("RicursiveForCausalLM") +@ModelBase.example("ricursiveAI/Ripple-65B-Preview") +class RicursiveModel(TextModel): + model_arch = gguf.MODEL_ARCH.RICURSIVE + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + hparams = self.hparams + + if (act := hparams.get("hidden_act", "silu")) != "silu": + raise ValueError(f"Ricursive: unsupported hidden_act {act!r}") + + # RicursiveConfig derives the layer pattern from layer_types alone, by default full attention iff i % 5 == 4 + if (layer_types := hparams.get("layer_types")) is None: + layer_types = ["full_attention" if i % 5 == 4 else "sliding_attention" for i in range(self.block_count)] + if len(layer_types) != self.block_count: + raise ValueError(f"Ricursive: layer_types has {len(layer_types)} entries, expected {self.block_count}") + if unknown := set(layer_types) - {"sliding_attention", "full_attention"}: + raise ValueError(f"Ricursive: unsupported layer_types {sorted(unknown)}") + self.is_swa = [t == "sliding_attention" for t in layer_types] + # only sliding layers use RoPE, an all-full-attention model would have no positional encoding + if not any(self.is_swa): + raise ValueError("Ricursive: layer_types has no sliding_attention layer, all-full-attention models are not supported") + # RicursiveConfig pops a stored hybrid_layer_pattern and recomputes it from layer_types + if (pattern := hparams.get("hybrid_layer_pattern")) is not None and list(pattern) != [int(swa) for swa in self.is_swa]: + logger.warning("Ricursive: hybrid_layer_pattern does not match layer_types, ignoring it") + + if not hparams.get("gate_full_precision", True): + logger.warning("Ricursive: gate_full_precision is false, llama.cpp computes the attention gate in f32") + + # RicursiveConfig defaults; the reference ignores the tensors of a disabled feature + self.use_valres = hparams.get("use_valres", False) + self.use_gate = hparams.get("use_gated_attention", True) + names = list(self.model_tensors) + if bias := [name for name in names if name.endswith("bias")]: + raise ValueError(f"Ricursive: bias tensors are not supported, found {bias[0]!r}") + for enabled, suffix in ((self.use_valres, ".self_attn.value_res_lambda"), (self.use_gate, ".self_attn.g_proj.weight")): + if enabled and sum(name.endswith(suffix) for name in names) != self.block_count: + raise ValueError(f"Ricursive: expected one {suffix!r} tensor per layer") + + def set_vocab(self): + self._set_vocab_gpt2() + # eos_token_id is [<|im_end|>, <|endoftext|>]: register the turn end as EOT so generation stops on it + eos = self.hparams.get("eos_token_id") + if isinstance(eos, list): + extra = [e for e in eos if e != self.hparams.get("bos_token_id")] + if extra: + self.gguf_writer.add_eot_token_id(extra[0]) + logger.info(f"gguf: eot_token_id = {extra[0]}") + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + self.gguf_writer.add_rope_dimension_count(hparams["head_dim"]) + + self.gguf_writer.add_sliding_window(hparams["sliding_window"]) + self.gguf_writer.add_sliding_window_pattern(self.is_swa) + # sliding layers use RoPE, full attention layers use no positional encoding + self.gguf_writer.add_rope_pattern(self.is_swa) + + if (embed_scale := hparams.get("embed_scale")) is not None: + if embed_scale == 0: + raise ValueError(f"Ricursive: unsupported embed_scale {embed_scale!r}") + self.gguf_writer.add_embedding_scale(float(embed_scale)) + # absent and 0 both mean no clamp (the reference tests `if clamp_threshold:`) + if (clamp := hparams.get("ffn_clamp_threshold")) is not None: + if clamp < 0 or not math.isfinite(clamp): + raise ValueError(f"Ricursive: unsupported ffn_clamp_threshold {clamp!r}") + self.gguf_writer.add_swiglu_clamp([float(clamp)] * self.block_count) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + # value_res_lambda has no .weight suffix in the checkpoint + if name.endswith(".value_res_lambda"): + if not self.use_valres: + return + name += ".weight" + elif name.endswith(".g_proj.weight") and not self.use_gate: + return + yield from super().modify_tensors(data_torch, name, bid) diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 9fe16a6fa..170df4e6f 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -162,6 +162,7 @@ class Keys: 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" + SWIGLU_CLAMP = "{arch}.swiglu_clamp" HIDDEN_ACT = "{arch}.hidden_activation" DENSE_FEAT_IN_SIZE = "{arch}.{dense}_feat_in" DENSE_FEAT_OUT_SIZE = "{arch}.{dense}_feat_out" @@ -710,6 +711,7 @@ class MODEL_ARCH(IntEnum): ARCEE = auto() AFMOE = auto() LAGUNA = auto() + RICURSIVE = auto() ERNIE4_5 = auto() ERNIE4_5_MOE = auto() HUNYUAN_MOE = auto() @@ -878,6 +880,7 @@ class MODEL_TENSOR(IntEnum): ATTN_K_NORM = auto() LAYER_OUT_NORM = auto() LAYER_OUT_SCALE = auto() + ATTN_VRES_LAMBDA = auto() PER_LAYER_TOKEN_EMBD = auto() # gemma3n PER_LAYER_MODEL_PROJ = auto() # gemma3n PER_LAYER_INP_GATE = auto() # gemma3n @@ -1624,6 +1627,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.ARCEE: "arcee", MODEL_ARCH.AFMOE: "afmoe", MODEL_ARCH.LAGUNA: "laguna", + MODEL_ARCH.RICURSIVE: "ricursive", MODEL_ARCH.ERNIE4_5: "ernie4_5", MODEL_ARCH.ERNIE4_5_MOE: "ernie4_5-moe", MODEL_ARCH.FALCON_H1: "falcon-h1", @@ -1791,6 +1795,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.FFN_POST_EXP_NORM: "blk.{bid}.ffn_post_exp_norm", MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm", MODEL_TENSOR.LAYER_OUT_SCALE: "blk.{bid}.layer_output_scale", + MODEL_TENSOR.ATTN_VRES_LAMBDA: "blk.{bid}.attn_vres_lambda", MODEL_TENSOR.PER_LAYER_TOKEN_EMBD: "per_layer_token_embd", # gemma3n MODEL_TENSOR.PER_LAYER_MODEL_PROJ: "per_layer_model_proj", # gemma3n MODEL_TENSOR.PER_LAYER_PROJ_NORM: "per_layer_proj_norm", # gemma3n @@ -5580,6 +5585,26 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_POST_NORM, MODEL_TENSOR.FFN_EXP_PROBS_B, ], + MODEL_ARCH.RICURSIVE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_GATE, + MODEL_TENSOR.ATTN_VRES_LAMBDA, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.FFN_PRE_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_POST_NORM, + ], MODEL_ARCH.LAGUNA: [ 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 0e7aa5012..f6b98a57d 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -912,6 +912,9 @@ class GGUFWriter: 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(self, values: Sequence[float]) -> None: + self.add_array(Keys.LLM.SWIGLU_CLAMP.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) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index ff22a60bc..eeadd6a9a 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -451,6 +451,10 @@ class TensorNameMap: "model.layers.{bid}.self_attn.attn_gate_proj", # k2-horizon ), + MODEL_TENSOR.ATTN_VRES_LAMBDA: ( + "model.layers.{bid}.self_attn.value_res_lambda", # ricursive + ), + # Feed-forward norm MODEL_TENSOR.FFN_NORM: ( "layers.{bid}.moe_connection.norm", # metis16 diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 6bb200f84..265c2ddeb 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -177,6 +177,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_K2_HORIZON, "k2-horizon" }, { LLM_ARCH_BEAM, "beam" }, { LLM_ARCH_KOLIBRI1, "kolibri1" }, + { LLM_ARCH_RICURSIVE, "ricursive" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -226,6 +227,7 @@ static const std::map LLM_KV_NAMES = { { 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_SWIGLU_CLAMP, "%s.swiglu_clamp" }, { 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" }, @@ -594,6 +596,7 @@ static const std::map LLM_TENSOR_NAMES = { { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" }, { LLM_TENSOR_LAYER_OUT_SCALE, "blk.%d.layer_output_scale" }, + { LLM_TENSOR_ATTN_VRES_LAMBDA, "blk.%d.attn_vres_lambda" }, { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, { LLM_TENSOR_POS_EMBD, "position_embd" }, { LLM_TENSOR_FFN_ACT, "blk.%d.ffn.act" }, @@ -1099,6 +1102,7 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_ATTN_K_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_LAYER_OUT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_LAYER_OUT_SCALE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_ATTN_VRES_LAMBDA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_ATTN_Q_A_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_ATTN_KV_A_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_ATTN_SUB_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, diff --git a/src/llama-arch.h b/src/llama-arch.h index 5dfad29ad..7bf0d3ae5 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -181,6 +181,7 @@ enum llm_arch { LLM_ARCH_BEAM, LLM_ARCH_KOLIBRI1, LLM_ARCH_LIMITE, + LLM_ARCH_RICURSIVE, LLM_ARCH_UNKNOWN, }; @@ -230,6 +231,7 @@ enum llm_kv { LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH, LLM_KV_SWIGLU_CLAMP_EXP, LLM_KV_SWIGLU_CLAMP_SHEXP, + LLM_KV_SWIGLU_CLAMP, LLM_KV_USE_PARALLEL_RESIDUAL, LLM_KV_TENSOR_DATA_LAYOUT, LLM_KV_EXPERT_COUNT, @@ -612,6 +614,7 @@ enum llm_tensor { LLM_TENSOR_ATTN_K_NORM, LLM_TENSOR_LAYER_OUT_NORM, LLM_TENSOR_LAYER_OUT_SCALE, + LLM_TENSOR_ATTN_VRES_LAMBDA, LLM_TENSOR_POST_ATTN_NORM, LLM_TENSOR_POST_MLP_NORM, LLM_TENSOR_PER_LAYER_TOKEN_EMBD, // gemma3n diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 95daf482d..2f7b97823 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -431,6 +431,7 @@ struct llama_hparams { // Step35: optional per-layer clamps for (Swi)GLU std::array swiglu_clamp_exp; // clamping for expert FFN std::array swiglu_clamp_shexp; // shared expert + std::array swiglu_clamp; // dense FFN // 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 diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp index b8a8f77df..b667bc49f 100644 --- a/src/llama-model-saver.cpp +++ b/src/llama-model-saver.cpp @@ -210,6 +210,8 @@ void llama_model_saver::add_kv_from_model() { add_kv(LLM_KV_MOE_LATENT_SIZE, hparams.moe_latent_size); add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp); add_kv(LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH, hparams.n_ff_chexp); + add_kv(LLM_KV_SWIGLU_CLAMP, std::vector( + hparams.swiglu_clamp.begin(), hparams.swiglu_clamp.begin() + hparams.n_layer_all)); add_kv(LLM_KV_SWIGLU_CLAMP_EXP, std::vector( hparams.swiglu_clamp_exp.begin(), hparams.swiglu_clamp_exp.begin() + hparams.n_layer_all)); add_kv(LLM_KV_SWIGLU_CLAMP_SHEXP, std::vector( diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 03aab320f..27809ccad 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -294,6 +294,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params return new llama_model_afmoe(params); case LLM_ARCH_LAGUNA: return new llama_model_laguna(params); + case LLM_ARCH_RICURSIVE: + return new llama_model_ricursive(params); case LLM_ARCH_ERNIE4_5: return new llama_model_ernie4_5(params); case LLM_ARCH_ERNIE4_5_MOE: @@ -1548,6 +1550,7 @@ void llama_model_base::load_hparams(llama_model_loader & ml) { 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); + std::fill(hparams.swiglu_clamp.begin(), hparams.swiglu_clamp.end(), 0.0f); ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff_arr, hparams.n_layer_all, false); ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer_all, false); @@ -3709,6 +3712,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_QWEN3MOE: case LLM_ARCH_LLADA_MOE: case LLM_ARCH_RND1: + case LLM_ARCH_RICURSIVE: case LLM_ARCH_OLMO2: case LLM_ARCH_OLMOE: case LLM_ARCH_PHI2: diff --git a/src/llama-model.h b/src/llama-model.h index 45d2b8666..f8cffb346 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -377,6 +377,9 @@ struct llama_layer { struct ggml_tensor * ffn_resid_alpha = nullptr; struct ggml_tensor * ffn_resid_beta = nullptr; + // value residual mix + struct ggml_tensor * attn_vres_lambda = nullptr; + // ff shared expert (shexp) struct ggml_tensor * ffn_gate_inp_shexp = nullptr; struct ggml_tensor * ffn_gate_shexp = nullptr; diff --git a/src/models/models.h b/src/models/models.h index 267468525..628275289 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -2269,6 +2269,18 @@ struct llama_model_beam : public llama_model_base { }; +struct llama_model_ricursive : public llama_model_base { + llama_model_ricursive(const struct llama_model_params & params) : llama_model_base(params) {} + void load_arch_hparams(llama_model_loader & ml) override; + void load_arch_tensors(llama_model_loader & ml) override; + + struct graph : public llm_graph_context { + graph(const llama_model & model, const llm_graph_params & params); + }; + + std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; +}; + struct llama_model_laguna : public llama_model_base { llama_model_laguna(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; diff --git a/src/models/ricursive.cpp b/src/models/ricursive.cpp new file mode 100644 index 000000000..cd5dc3547 --- /dev/null +++ b/src/models/ricursive.cpp @@ -0,0 +1,236 @@ +#include "models.h" + +void llama_model_ricursive::load_arch_hparams(llama_model_loader & ml) { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale, false); + // optional: absent or 0 means no clamp (the reference tests `if clamp_threshold:`); a scalar applies to every layer + ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP, hparams.swiglu_clamp, hparams.n_layer_all, false); + for (uint32_t il = 0; il < hparams.n_layer_all; ++il) { + if (!std::isfinite(hparams.swiglu_clamp[il]) || hparams.swiglu_clamp[il] < 0.0f) { + throw std::runtime_error(format("ricursive: swiglu_clamp for layer %u must be finite and non-negative, got %g", il, hparams.swiglu_clamp[il])); + } + } + + if (hparams.n_embd_head_k() != hparams.n_embd_head_v()) { + throw std::runtime_error("ricursive: key_length and value_length must be equal"); + } + + // RoPE is only applied on the SWA layers, so a model without SWA would have no positional encoding + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); + if (hparams.n_swa == 0) { + throw std::runtime_error("ricursive: sliding_window must be > 0"); + } + hparams.swa_type = LLAMA_SWA_TYPE_STANDARD; + if (ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, false)) { + // load_swa_pattern only bounds the array by LLAMA_MAX_LAYERS: a short one would leave the last layers dense and without RoPE + uint32_t n_entries = 0; + ml.get_arr_n(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, n_entries); + if (n_entries != hparams.n_layer()) { + throw std::runtime_error(format("ricursive: sliding_window_pattern has %u entries, expected %u", n_entries, hparams.n_layer())); + } + } else { + // default: every 5th layer is global (RicursiveConfig layer_types: full attention iff i % 5 == 4) + load_swa_pattern(ml, 5); + } + if (!hparams.is_swa_any()) { + throw std::runtime_error("ricursive: no sliding window layer, the model would have no positional encoding"); + } + + // RoPE exactly on the SWA layers; rope_pattern is optional because the first published GGUFs do not carry it + hparams.rope_pattern = hparams.is_swa_impl; + ml.get_key_or_arr(LLM_KV_ATTENTION_ROPE_PATTERN, hparams.rope_pattern, hparams.n_layer(), false); + for (uint32_t il = 0; il < hparams.n_layer(); ++il) { + if (hparams.has_rope(il) != hparams.is_swa(il)) { + throw std::runtime_error(format("ricursive: layer %u has %s but %s", il, + hparams.has_rope(il) ? "RoPE" : "no RoPE", hparams.is_swa(il) ? "a sliding window" : "no sliding window")); + } + } + + // the SWA layers are the only RoPE layers, they use the model RoPE parameters + hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train; + hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train; + ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false); + + switch (hparams.n_layer()) { + case 80: type = LLM_TYPE_65B; break; + default: type = LLM_TYPE_UNKNOWN; + } +} + +void llama_model_ricursive::load_arch_tensors(llama_model_loader &) { + LLAMA_LOAD_LOCALS; + + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + 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 == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + const int64_t n_embd_q = n_embd_head_k * n_head; + + 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.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_q}, 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_q, n_embd}, 0); + + 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); + + layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_embd_q}, TENSOR_NOT_REQUIRED); + layer.attn_vres_lambda = create_tensor(tn(LLM_TENSOR_ATTN_VRES_LAMBDA, "weight", i), {1}, TENSOR_NOT_REQUIRED); + + layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "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_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0); + } + + // the gate and the value residual are per-model switches (use_gated_attention, use_valres) + for (int i = 1; i < n_layer; ++i) { + if ((layers[i].wqkv_gate == nullptr) != (layers[0].wqkv_gate == nullptr)) { + throw std::runtime_error(format("ricursive: attn_gate must be present in all layers or in none (layers 0 and %d differ)", i)); + } + if ((layers[i].attn_vres_lambda == nullptr) != (layers[0].attn_vres_lambda == nullptr)) { + throw std::runtime_error(format("ricursive: attn_vres_lambda must be present in all layers or in none (layers 0 and %d differ)", i)); + } + } +} + +std::unique_ptr llama_model_ricursive::build_arch_graph(const llm_graph_params & params) const { + return std::make_unique(params); +} + +llama_model_ricursive::graph::graph(const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v(); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // value residual: x0 = token embeddings * embedding_scale, the same tensor for all layers (modeling_ricursive.py:380-381,405) + ggml_tensor * inp_vres = inpL; + + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_iswa(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); + + for (int il = 0; il < n_layer; ++il) { + const auto & layer = model.layers[il]; + + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, layer.attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + ggml_tensor * attn_in = cur; + + { + ggml_tensor * Qcur = build_lora_mm(layer.wq, attn_in, layer.wq_s); + ggml_tensor * Kcur = build_lora_mm(layer.wk, attn_in, layer.wk_s); + + // v_in = lerp(x0, attn_norm(x), sigmoid(lambda)), lambda is learned per layer (modeling_ricursive.py:182-185) + ggml_tensor * v_in = attn_in; + if (layer.attn_vres_lambda) { + ggml_tensor * mix = ggml_sigmoid(ctx0, layer.attn_vres_lambda); + v_in = ggml_add(ctx0, inp_vres, ggml_mul(ctx0, ggml_sub(ctx0, attn_in, inp_vres), mix)); + cb(v_in, "vres_in", il); + } + ggml_tensor * Vcur = build_lora_mm(layer.wv, v_in, layer.wv_s); + + 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); + + Qcur = build_norm(Qcur, layer.attn_q_norm, NULL, LLM_NORM_RMS, il); + Kcur = build_norm(Kcur, layer.attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + cb(Kcur, "Kcur_normed", il); + + // global layers use no RoPE (NoPE): the reference rotates only sliding_attention layers (modeling_ricursive.py:127-129,203-205) + if (hparams.has_rope(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); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, + n_rot, 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, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + } + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, NULL, NULL, NULL, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + + if (layer.wqkv_gate) { + ggml_tensor * gate = build_lora_mm(layer.wqkv_gate, attn_in, layer.wqkv_gate_s); + gate = ggml_sigmoid(ctx0, gate); + cb(gate, "attn_gate", il); + cur = ggml_mul(ctx0, cur, gate); + cb(cur, "attn_gated", il); + } + + cur = build_lora_mm(layer.wo, cur, layer.wo_s); + } + + 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); + } + + cur = build_norm(cur, layer.attn_post_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, layer.ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + { + ggml_tensor * gate = build_lora_mm(layer.ffn_gate, cur, layer.ffn_gate_s); + ggml_tensor * up = build_lora_mm(layer.ffn_up, cur, layer.ffn_up_s); + cur = hparams.swiglu_clamp[il] > 0.0f + ? ggml_swiglu_clamp(ctx0, gate, up, hparams.swiglu_clamp[il]) + : ggml_swiglu_split(ctx0, gate, up); + cb(cur, "ffn_swiglu", il); + cur = build_lora_mm(layer.ffn_down, cur, layer.ffn_down_s); + cb(cur, "ffn_out", il); + } + + cur = build_norm(cur, layer.ffn_post_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_post_norm", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + inpL = cur; + } + + cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1); + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur, model.output_s); + cb(cur, "result_output", -1); + res->t_logits = cur; +} diff --git a/tests/fusion/MTL.csv b/tests/fusion/MTL.csv index cae85fc21..46aec1076 100644 --- a/tests/fusion/MTL.csv +++ b/tests/fusion/MTL.csv @@ -318,6 +318,8 @@ qwen4exp ,0 ,any ,SOFT_MAX+ARGSORT+GET_ROWS+SUM_ROWS+CLAMP+DIV, qwen4exp ,0 ,any ,SSM_CONV+UNARY , 1 refact ,0 ,any ,RMS_NORM+MUL , 5 refact ,0 ,any ,RMS_NORM+MUL , 5 +ricursive ,0 ,any ,RMS_NORM+MUL , 9 +ricursive ,0 ,any ,RMS_NORM+MUL+ADD , 4 rnd1 ,0 ,any ,MUL+ADD , 4 rnd1 ,0 ,any ,RMS_NORM+MUL , 9 rnd1 ,0 ,any ,SOFT_MAX+ARGSORT+GET_ROWS+SUM_ROWS+CLAMP+DIV, 2 diff --git a/tests/test-llama-archs.cpp b/tests/test-llama-archs.cpp index c42831d0e..39255c341 100644 --- a/tests/test-llama-archs.cpp +++ b/tests/test-llama-archs.cpp @@ -267,7 +267,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) { n_layer = 6; } else if (arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_LAGUNA) { n_embd = 160; // exercise per-head tensor split granularity with head size 80 - } else if (arch == LLM_ARCH_QWEN3 || arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_AFMOE) { + } else if (arch == LLM_ARCH_QWEN3 || arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_AFMOE || arch == LLM_ARCH_RICURSIVE) { n_head = 4; } else if (arch == LLM_ARCH_LIMITE) { n_head = 4; @@ -311,7 +311,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) { if (arch == LLM_ARCH_QWEN3) { n_head_kv = 1; // MQA coverage } else if (arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_MUSE_SPARK || arch == LLM_ARCH_AFMOE || arch == LLM_ARCH_MAION_CODER || - arch == LLM_ARCH_KOLIBRI1 || arch == LLM_ARCH_LIMITE || arch == LLM_ARCH_GEMMA4) { + arch == LLM_ARCH_KOLIBRI1 || arch == LLM_ARCH_LIMITE || arch == LLM_ARCH_GEMMA4 || arch == LLM_ARCH_RICURSIVE) { n_head_kv = 2; // GQA coverage } if (arch == LLM_ARCH_BEAM) { @@ -466,7 +466,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) { ms.add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, pattern); } else if (arch == LLM_ARCH_COHERE2MOE || arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_SPARK2_5 || arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_MUSE_SPARK || arch == LLM_ARCH_GRANITE_SWA || arch == LLM_ARCH_DOTS3NOTE || - arch == LLM_ARCH_KOLIBRI1 || arch == LLM_ARCH_MAPLE || arch == LLM_ARCH_FLINT || arch == LLM_ARCH_BEAM) { + arch == LLM_ARCH_KOLIBRI1 || arch == LLM_ARCH_MAPLE || arch == LLM_ARCH_FLINT || arch == LLM_ARCH_BEAM || arch == LLM_ARCH_RICURSIVE) { std::vector pattern; pattern.reserve(n_layer); for (uint32_t il = 0; il < n_layer; il++) { @@ -636,6 +636,17 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) { ms.add_kv(LLM_KV_SWIGLU_CLAMP_EXP, 7.0f); } + if (arch == LLM_ARCH_RICURSIVE) { + std::vector swiglu_clamp; + swiglu_clamp.reserve(n_layer); + for (uint32_t il = 0; il < n_layer; il++) { + // the random weight scales keep the FFN activations small, a small limit still clamps so the roundtrip checks the per-layer values + swiglu_clamp.push_back(il % 3 == 1 ? 0.0f : 1e-3f); + } + ms.add_kv(LLM_KV_EMBEDDING_SCALE, 16.0f); + ms.add_kv(LLM_KV_SWIGLU_CLAMP, swiglu_clamp); + } + if (arch == LLM_ARCH_METIS16) { ms.add_kv(LLM_KV_METIS_PASS_COUNT, uint32_t(2)); ms.add_kv(LLM_KV_HYPER_CONNECTION_COUNT, uint32_t(4));