upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-10-02 19:07:25 -05:00
Code Issues Packages Projects Releases Wiki Activity
Files
b11316
llama.cpp/src/llama-arch.h
T
timkhronosSigbjørn SkjæretStanisław SzymczykPiotr Wilkin
 649dcb1036 add GLM-5.3-Flash (GLM5-Next) support (#27773) 
* Rebase GLM-Next support onto master, and migrate to llama-memory-hybrid-idx

* Add initial MTP support

* Merge branch optimizations. Reduce allocated compute buffer size, speed up long context decode, fla, and slight MTP improvements.

* Review driven changes, remove env vars, protect tensors

* Strip MTP for initial PR

* Clean up after mtp strip

* Clean up after mtp strip

* Update speculative.cpp

* Update llama-context.h

* Clean up after mtp strip

* Fix tokenizer ignore merges

* Improve quantization protection selection

* Refactor mhc helpers, graph base

* Lint Fixes

* Apply suggestions from code review

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Skip glm5-next in model saver, fix CRLF

* Skip glm5-next in sweep

* Remove T4 fallback

* Review cleanup

* Review suggestions

* Defer separate MTP gguf handling to MTP PR, drop filter

* Repad n_head_kv

* kpool init apply

* Order by descending score

* Drop guard

* read kpool from hparams, clarify kpool cache flags, remove kpool_build_state(nullptr)

* Add glm5-next support to model saver and add arch test fixture

* Review cleanup

* Kpool pooled caching clarify

* Add multi stream support

* Finish Rebase

* Sparse FA fir DSA prefill

* Const

* Update llama-model.cpp to fix rebase error

* gguf-py : merge tensor map entries for HC tensors

* model : use build_gdn_l2_norm in GLM5_NEXT implementation

* chore : remove trailing whitespace

* model : use new OP precision setting API in GLM5_NEXT implementation

* mtmd : use ggml_swiglu_clamp in GLM5V and apply the image token limit

The two clamps around swiglu_split are what ggml_swiglu_clamp already does,
so the clamp bounds collapse back to one value. GLM5V also never called
set_limit_image_tokens(), so --image-max-tokens had no effect.

Assisted-by: Claude Opus 5
(cherry picked from commit 46d18e12d422be4cc04a70e4a9a9e0168bb3d5b7)

* llama : keep the GLM5-Next k-pool layout across ubatches

The layout was rebuilt from a full cell scan on every ubatch. Pools are fixed
by the positions relative to the sequence's first one, so the layout now lives
on the memory and a ubatch only appends to it.

A sequence edit no longer stales every pooled key either, only the ones at or
after the edited position, which makes a tail seq_rm free. The pooling subgraph
is built unconditionally so the graph shape no longer changes every kpool
tokens, and the pool axis is folded into rows before soft_max, which otherwise
exceeds the CUDA gridDim.y limit past n_kv 262144.

Assisted-by: Claude Opus 5
(cherry picked from commit 5d1c40b93e17fddbf73b785efe43e0d02ccb3977)

* model : write the GLM5-Next recurrent rollback checkpoints

The conv state and the delta net state were only written to the live row, so a
rollback restored whatever the checkpoint rows happened to hold. Take the same
route as kimi-k3: build_recurrent_attn for the state, and write all K_rs conv
groups. That also drops a state view that assumed contiguous rows.

Enroll the arch in test-recurrent-state-rollback, which catches this under its
garbage-filled cache pass.

Assisted-by: Claude Opus 5
(cherry picked from commit 5ace37e86d5d448e83ef5dde5632c748185b18cd)

* llama: fix PR #27773 test-save-load-state restore failure

Clear the attention and indexer cache data after a failed hybrid state restore so restored NaNs cannot affect a later sequence.

Assisted-by: Codex

* llama: fix PR #27773 gpu-rocm graph reallocation

Reserve the full GLM5-Next pool capacity and dirty pool count. The gpu-rocm Test step aborts when n_new grows while the graph node count stays fixed; CUDA, Vulkan, Metal, and WebGPU checks report the same error.

Assisted-by: Codex

* llama : fix GLM5-Next k-pool layout staleness after edits and shared teardown

Two defects in the cross-ubatch k-pool layout added by the k-pool commit:

1. Wrong results. An edited sequence only rebuilt its pool layout when its cell
   count changed, so if the first ubatch after an edit added back exactly as many
   cells as were removed, the stale position-to-cell list survived. With a unified
   cache and more than one sequence, where another sequence takes the freed cells,
   the reused layout points at the wrong cells (CPU: large logit drift, CUDA: NaN).
   Rebuild whenever the sequence is stale, not only on a size mismatch.

2. Slowdown. "shared" mode was assumed to end only with an edit that forces a
   rebuild, but sharing also ends when the other sequence is removed. The survivor
   kept shared = true, pinning cache_safe off and re-pooling every pool on every
   ubatch (server trigger: n>1 completions with -kvu, via the seq_cp in
   copy_state_to). In seq_rm, if the layout has shared cells, stale every sequence
   so one rebuild re-derives sharing and cache_safe returns to 1.

Assisted-by: Claude Opus 5

* llama : fix build_attn_mha stream stride for non-contiguous q

build_attn_mha split the batch into streams with a stream stride of
q->nb[3]/n_stream. That only equals one stream's span, (ne[2]/n_stream)*nb[2],
when q is contiguous. GLM5-Next is nope-only, so it does not concat a rope part
and passes the permuted q_absorbed straight in, where nb[3] != ne[2]*nb[2]; the
stride was then n_head times too large and every stream s >= 1 read another
head's queries. Split-KV (-np N without --kv-unified) multi-stream prefill was
wrong for every stream past the first. Unified KV and decode were unaffected
(n_stream == 1, and decode takes the gather path). Other MLA models concat rope
so q is contiguous and the computed value is unchanged for them.

Compute the stride from the token dimension, which is identical for a
contiguous q.

Assisted-by: Claude Opus 5

* llama : re-derive GLM5-Next k-pool sharing on state_read/state_drop

The shared-cell teardown added to seq_rm (stale every sequence when the layout
has shared cells, so a survivor does not keep shared = true and pin cache_safe
off) was missing from the other paths that can free shared cells: state_read
and state_drop staled only the one sequence. Apply the same re-derivation there
and correct the comment that claimed sharing ends only via an edit or seq_rm.

Assisted-by: Claude Opus 5

* quant : drop duplicate GLM5-Next hc_ filter

The hc_ name filter was listed twice in the GLM5_NEXT protection block.

Assisted-by: Claude Opus 5

* glm5-next: scope K-pool cache access to indexed operations

* glm5-next: keep K-pool access in hybrid index memory

* glm5-next: keep mHC graph builders model-local

* glm5-next: mark only touched pools per ubatch

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
Co-authored-by: Piotr Wilkin <ilintar@gmail.com>
2026-09-30 14:20:32 +08:00

806 lines
23 KiB
C++
Raw Permalink Blame History

#pragma once
#include "ggml.h" // ggml_op
#include <string>
#include <set>
#include <vector>
//
// gguf constants (sync with gguf.py)
//
enum llm_arch {
LLM_ARCH_CLIP,
LLM_ARCH_LLAMA,
LLM_ARCH_LLAMA4,
LLM_ARCH_DECI,
LLM_ARCH_FALCON,
LLM_ARCH_BAICHUAN,
LLM_ARCH_GROK,
LLM_ARCH_GPT2,
LLM_ARCH_GPTJ,
LLM_ARCH_GPTNEOX,
LLM_ARCH_MPT,
LLM_ARCH_STARCODER,
LLM_ARCH_REFACT,
LLM_ARCH_BERT,
LLM_ARCH_MODERN_BERT,
LLM_ARCH_NOMIC_BERT,
LLM_ARCH_NOMIC_BERT_MOE,
LLM_ARCH_NEO_BERT,
LLM_ARCH_JINA_BERT_V2,
LLM_ARCH_JINA_BERT_V3,
LLM_ARCH_EUROBERT,
LLM_ARCH_BLOOM,
LLM_ARCH_STABLELM,
LLM_ARCH_QWEN,
LLM_ARCH_QWEN2,
LLM_ARCH_QWEN2MOE,
LLM_ARCH_QWEN2VL,
LLM_ARCH_QWEN3,
LLM_ARCH_QWEN3MOE,
LLM_ARCH_QWEN3NEXT,
LLM_ARCH_QWEN3VL,
LLM_ARCH_QWEN3VLMOE,
LLM_ARCH_QWEN35,
LLM_ARCH_QWEN35MOE,
LLM_ARCH_QWEN4EXP,
LLM_ARCH_PHI2,
LLM_ARCH_PHI3,
LLM_ARCH_PHIMOE,
LLM_ARCH_PLAMO,
LLM_ARCH_PLAMO2,
LLM_ARCH_PLAMO3,
LLM_ARCH_CODESHELL,
LLM_ARCH_ORION,
LLM_ARCH_INTERNLM2,
LLM_ARCH_MINICPM,
LLM_ARCH_MINICPM3,
LLM_ARCH_GEMMA,
LLM_ARCH_GEMMA2,
LLM_ARCH_GEMMA3,
LLM_ARCH_GEMMA3N,
LLM_ARCH_GEMMA4,
LLM_ARCH_GEMMA4_ASSISTANT,
LLM_ARCH_GEMMA_EMBEDDING,
LLM_ARCH_STARCODER2,
LLM_ARCH_MAMBA,
LLM_ARCH_MAMBA2,
LLM_ARCH_MAPLE,
LLM_ARCH_JAMBA,
LLM_ARCH_FALCON_H1,
LLM_ARCH_XVERSE,
LLM_ARCH_COMMAND_R,
LLM_ARCH_COHERE2,
LLM_ARCH_COHERE2MOE,
LLM_ARCH_DBRX,
LLM_ARCH_OLMO,
LLM_ARCH_OLMO2,
LLM_ARCH_OLMOE,
LLM_ARCH_MUSE_GLIMMER,
LLM_ARCH_OPENELM,
LLM_ARCH_ARCTIC,
LLM_ARCH_DEEPSEEK,
LLM_ARCH_DEEPSEEK2,
LLM_ARCH_DEEPSEEK2OCR,
LLM_ARCH_DEEPSEEK32,
LLM_ARCH_DEEPSEEK4,
LLM_ARCH_CHATGLM,
LLM_ARCH_GLM4,
LLM_ARCH_GLM4_MOE,
LLM_ARCH_GLM_DSA,
LLM_ARCH_BITNET,
LLM_ARCH_T5,
LLM_ARCH_T5ENCODER,
LLM_ARCH_JAIS,
LLM_ARCH_JAIS2,
LLM_ARCH_NEMOTRON,
LLM_ARCH_NEMOTRON_H,
LLM_ARCH_NEMOTRON_H_MOE,
LLM_ARCH_EXAONE,
LLM_ARCH_EXAONE4,
LLM_ARCH_EXAONE_MOE,
LLM_ARCH_RWKV6,
LLM_ARCH_RWKV6QWEN2,
LLM_ARCH_RWKV7,
LLM_ARCH_ARWKV7,
LLM_ARCH_GRANITE,
LLM_ARCH_GRANITE_MOE,
LLM_ARCH_GRANITE_HYBRID,
LLM_ARCH_GRANITE_SWITCH,
LLM_ARCH_GRANITE_SWA,
LLM_ARCH_CHAMELEON,
LLM_ARCH_WAVTOKENIZER_DEC,
LLM_ARCH_PLM,
LLM_ARCH_BAILINGMOE,
LLM_ARCH_BAILINGMOE2,
LLM_ARCH_BAILINGMOE3,
LLM_ARCH_DOTS1,
LLM_ARCH_DOTS3NOTE,
LLM_ARCH_ARCEE,
LLM_ARCH_AFMOE,
LLM_ARCH_LAGUNA,
LLM_ARCH_ERNIE4_5,
LLM_ARCH_ERNIE4_5_MOE,
LLM_ARCH_HUNYUAN_MOE,
LLM_ARCH_HUNYUAN_DENSE,
LLM_ARCH_HUNYUAN_VL,
LLM_ARCH_HY_V3,
LLM_ARCH_HY_V4,
LLM_ARCH_SMOLLM3,
LLM_ARCH_OPENAI_MOE,
LLM_ARCH_LFM2,
LLM_ARCH_LFM2MOE,
LLM_ARCH_DREAM,
LLM_ARCH_SMALLTHINKER,
LLM_ARCH_LLADA,
LLM_ARCH_LLADA_MOE,
LLM_ARCH_SEED_OSS,
LLM_ARCH_GROVEMOE,
LLM_ARCH_APERTUS,
LLM_ARCH_MINIMAX_M2,
LLM_ARCH_COGVLM,
LLM_ARCH_RND1,
LLM_ARCH_PANGU_EMBED,
LLM_ARCH_MISTRAL3,
LLM_ARCH_MISTRAL4,
LLM_ARCH_PADDLEOCR,
LLM_ARCH_MIMO2,
LLM_ARCH_STEP35,
LLM_ARCH_SPARK2_5,
LLM_ARCH_LLAMA_EMBED,
LLM_ARCH_MAINCODER,
LLM_ARCH_KIMI_LINEAR,
LLM_ARCH_KIMI_K3,
LLM_ARCH_GLM5_NEXT,
LLM_ARCH_TALKIE,
LLM_ARCH_MELLUM,
LLM_ARCH_EAGLE3,
LLM_ARCH_MINIMAX_M3,
LLM_ARCH_DFLASH,
LLM_ARCH_NANBEIGE,
LLM_ARCH_QWEN3TTS,
LLM_ARCH_POCKETTTS,
LLM_ARCH_MINIMAX_01,
LLM_ARCH_HRM_TEXT,
LLM_ARCH_UNKNOWN,
};
enum llm_kv {
LLM_KV_GENERAL_TYPE,
LLM_KV_GENERAL_ARCHITECTURE,
LLM_KV_GENERAL_QUANTIZATION_VERSION,
LLM_KV_GENERAL_ALIGNMENT,
LLM_KV_GENERAL_FILE_TYPE,
LLM_KV_GENERAL_SAMPLING_SEQUENCE,
LLM_KV_GENERAL_SAMPLING_TOP_K,
LLM_KV_GENERAL_SAMPLING_TOP_P,
LLM_KV_GENERAL_SAMPLING_MIN_P,
LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY,
LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD,
LLM_KV_GENERAL_SAMPLING_TEMP,
LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N,
LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT,
LLM_KV_GENERAL_SAMPLING_MIROSTAT,
LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU,
LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA,
LLM_KV_GENERAL_NAME,
LLM_KV_GENERAL_AUTHOR,
LLM_KV_GENERAL_VERSION,
LLM_KV_GENERAL_URL,
LLM_KV_GENERAL_DESCRIPTION,
LLM_KV_GENERAL_LICENSE,
LLM_KV_GENERAL_SOURCE_URL,
LLM_KV_GENERAL_SOURCE_HF_REPO,
LLM_KV_GENERAL_TENSOR_EXTRA_NAME,
LLM_KV_GENERAL_TENSOR_EXTRA_PREC_A4,
LLM_KV_VOCAB_SIZE,
LLM_KV_CONTEXT_LENGTH,
LLM_KV_EMBEDDING_LENGTH,
LLM_KV_EMBEDDING_LENGTH_OUT,
LLM_KV_EMBEDDING_LENGTH_PER_LAYER,
LLM_KV_FEATURES_LENGTH,
LLM_KV_BLOCK_COUNT,
LLM_KV_LEADING_DENSE_BLOCK_COUNT,
LLM_KV_ATTN_RES_BLOCK_SIZE,
LLM_KV_ACTIVATION_SITU_BETA,
LLM_KV_ACTIVATION_SITU_LINEAR_BETA,
LLM_KV_FEED_FORWARD_LENGTH,
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,
LLM_KV_EXPERT_USED_COUNT,
LLM_KV_EXPERT_SHARED_COUNT,
LLM_KV_EXPERT_GROUP_COUNT,
LLM_KV_EXPERT_GROUP_USED_COUNT,
LLM_KV_EXPERT_WEIGHTS_SCALE,
LLM_KV_EXPERT_WEIGHTS_NORM,
LLM_KV_EXPERT_LATENT_LENGTH,
LLM_KV_EXPERT_GATING_FUNC,
LLM_KV_EXPERT_GROUP_SCALE,
LLM_KV_EXPERTS_PER_GROUP,
LLM_KV_MOE_EVERY_N_LAYERS,
LLM_KV_MOE_LATENT_SIZE,
LLM_KV_NEXTN_PREDICT_LAYERS,
LLM_KV_NUM_DEEPSTACK_LAYERS,
LLM_KV_DEEPSTACK_MAPPING,
LLM_KV_HIDDEN_ACT,
LLM_KV_POOLING_TYPE,
LLM_KV_LOGIT_SCALE,
LLM_KV_DECODER_START_TOKEN_ID,
LLM_KV_DECODER_BLOCK_COUNT,
LLM_KV_ATTN_LOGIT_SOFTCAPPING,
LLM_KV_ROUTER_LOGIT_SOFTCAPPING,
LLM_KV_FINAL_LOGIT_SOFTCAPPING,
LLM_KV_SWIN_NORM,
LLM_KV_RESCALE_EVERY_N_LAYERS,
LLM_KV_TIME_MIX_EXTRA_DIM,
LLM_KV_TIME_DECAY_EXTRA_DIM,
LLM_KV_RESIDUAL_SCALE,
LLM_KV_EMBEDDING_SCALE,
LLM_KV_ADAPTER_COUNT,
LLM_KV_ADAPTER_TOKEN_IDS_ACTIVATE,
LLM_KV_ADAPTER_TOKEN_IDS_SUBSTITUTE,
LLM_KV_ADAPTER_LORA_RANK,
LLM_KV_ADAPTER_ROUTER_GAIN,
LLM_KV_TOKEN_SHIFT_COUNT,
LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
LLM_KV_FULL_ATTENTION_INTERVAL,
LLM_KV_NUM_LOOPS,
LLM_KV_SKIP_LOOP_FINAL_NORM,
LLM_KV_HRM_LAYERS_PER_STACK,
LLM_KV_HRM_H_CYCLES,
LLM_KV_HRM_L_CYCLES,
LLM_KV_HRM_PREFIX_LM,
LLM_KV_ATTENTION_HEAD_COUNT,
LLM_KV_ATTENTION_HEAD_COUNT_KV,
LLM_KV_ATTENTION_MAX_ALIBI_BIAS,
LLM_KV_ATTENTION_CLAMP_KQV,
LLM_KV_ATTENTION_KEY_LENGTH,
LLM_KV_ATTENTION_VALUE_LENGTH,
LLM_KV_ATTENTION_LAYERNORM_EPS,
LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,
LLM_KV_ATTENTION_GROUPNORM_EPS,
LLM_KV_ATTENTION_GROUPNORM_GROUPS,
LLM_KV_ATTENTION_CAUSAL,
LLM_KV_ATTENTION_Q_LORA_RANK,
LLM_KV_ATTENTION_KV_LORA_RANK,
LLM_KV_ATTENTION_DECAY_LORA_RANK,
LLM_KV_ATTENTION_ICLR_LORA_RANK,
LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK,
LLM_KV_ATTENTION_GATE_LORA_RANK,
LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
LLM_KV_ATTENTION_SLIDING_WINDOW,
LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,
LLM_KV_ATTENTION_SCALE,
LLM_KV_ATTENTION_ROPE_PATTERN,
LLM_KV_ATTENTION_OUTPUT_SCALE,
LLM_KV_ATTENTION_VALUE_SCALE,
LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
LLM_KV_ATTENTION_TEMPERATURE_SCALE,
LLM_KV_ATTENTION_KEY_LENGTH_MLA,
LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
LLM_KV_ATTENTION_KEY_LENGTH_SWA,
LLM_KV_ATTENTION_VALUE_LENGTH_SWA,
LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA,
LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA,
LLM_KV_ATTENTION_KV_LORA_RANK_SWA,
LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,
LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,
LLM_KV_ATTENTION_INDEXER_TOP_K,
LLM_KV_ATTENTION_INDEXER_BLOCK_SIZE,
LLM_KV_ATTENTION_INDEXER_LOCAL_BLOCKS,
LLM_KV_ATTENTION_INDEXER_TYPES,
LLM_KV_ATTENTION_INDEXER_KPOOL,
LLM_KV_ATTENTION_INDEXER_KPOOL_SELECT_TAIL,
LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT,
LLM_KV_ATTENTION_OUTPUT_LORA_RANK,
LLM_KV_ATTENTION_COMPRESS_ROPE_FREQ_BASE,
LLM_KV_ATTENTION_COMPRESS_RATIOS,
LLM_KV_ATTENTION_SHARED_KV_LAYERS,
LLM_KV_ATTENTION_RECURRENT_LAYERS,
LLM_KV_HYPER_CONNECTION_COUNT,
LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS,
LLM_KV_HYPER_CONNECTION_EPSILON,
LLM_KV_HYPER_CONNECTION_MAGNITUDE,
LLM_KV_HYPER_CONNECTION_LOW_RANK,
LLM_KV_PLE_LAYERS,
LLM_KV_PLE_NGRAM_SIZE,
LLM_KV_PLE_HEADS_PER_NGRAM,
LLM_KV_PLE_CONV_KERNEL,
LLM_KV_PLE_LAYER_MULTIPLIERS,
LLM_KV_PLE_HEAD_OFFSETS,
LLM_KV_PLE_HEAD_VOCAB_SIZES,
LLM_KV_PLE_EOS_TOKEN_ID,
LLM_KV_PLE_IMAGE_TOKEN_ID,
LLM_KV_HASH_LAYER_COUNT,
LLM_KV_ROPE_DIMENSION_COUNT,
LLM_KV_ROPE_DIMENSION_COUNT_SWA,
LLM_KV_ROPE_DIMENSION_SECTIONS,
LLM_KV_ROPE_FREQ_BASE,
LLM_KV_ROPE_FREQ_BASE_SWA,
LLM_KV_ROPE_SCALE_LINEAR,
LLM_KV_ROPE_SCALING_TYPE,
LLM_KV_ROPE_SCALING_FACTOR,
LLM_KV_ROPE_SCALING_ALPHA,
LLM_KV_ROPE_SCALING_ATTN_FACTOR,
LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,
LLM_KV_ROPE_SCALING_FINETUNED,
LLM_KV_ROPE_SCALING_YARN_LOG_MUL,
LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR,
LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR,
LLM_KV_ROPE_SCALING_YARN_BETA_FAST,
LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,
LLM_KV_SPLIT_NO,
LLM_KV_SPLIT_COUNT,
LLM_KV_SPLIT_TENSORS_COUNT,
LLM_KV_SSM_INNER_SIZE,
LLM_KV_SSM_CONV_KERNEL,
LLM_KV_SSM_STATE_SIZE,
LLM_KV_SSM_TIME_STEP_RANK,
LLM_KV_SSM_GROUP_COUNT,
LLM_KV_SSM_DT_B_C_RMS,
LLM_KV_KDA_HEAD_DIM,
LLM_KV_KDA_SAFE_GATE,
LLM_KV_KDA_GATE_LOWER_BOUND,
LLM_KV_WKV_HEAD_SIZE,
LLM_KV_TOKENIZER_MODEL,
LLM_KV_TOKENIZER_PRE,
LLM_KV_TOKENIZER_LIST,
LLM_KV_TOKENIZER_TOKEN_TYPE,
LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,
LLM_KV_TOKENIZER_SCORES,
LLM_KV_TOKENIZER_MERGES,
LLM_KV_TOKENIZER_BOS_ID,
LLM_KV_TOKENIZER_EOS_ID,
LLM_KV_TOKENIZER_EOT_ID,
LLM_KV_TOKENIZER_EOM_ID,
LLM_KV_TOKENIZER_UNK_ID,
LLM_KV_TOKENIZER_SEP_ID,
LLM_KV_TOKENIZER_PAD_ID,
LLM_KV_TOKENIZER_CLS_ID,
LLM_KV_TOKENIZER_MASK_ID,
LLM_KV_TOKENIZER_ADD_BOS,
LLM_KV_TOKENIZER_ADD_EOS,
LLM_KV_TOKENIZER_ADD_SEP,
LLM_KV_TOKENIZER_ADD_PREFIX,
LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,
LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
LLM_KV_TOKENIZER_HF_JSON,
LLM_KV_TOKENIZER_RWKV,
LLM_KV_TOKENIZER_CHAT_TEMPLATE,
LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE,
LLM_KV_TOKENIZER_NORMALIZER_STRIP_ACCENTS,
LLM_KV_TOKENIZER_FIM_PRE_ID,
LLM_KV_TOKENIZER_FIM_SUF_ID,
LLM_KV_TOKENIZER_FIM_MID_ID,
LLM_KV_TOKENIZER_FIM_PAD_ID,
LLM_KV_TOKENIZER_FIM_REP_ID,
LLM_KV_TOKENIZER_FIM_SEP_ID,
LLM_KV_TOKENIZER_SUPPRESS_TOKENS,
LLM_KV_ADAPTER_TYPE,
LLM_KV_ADAPTER_LORA_ALPHA,
LLM_KV_ADAPTER_LORA_TASK_NAME,
LLM_KV_ADAPTER_LORA_PROMPT_PREFIX,
LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS,
LLM_KV_POSNET_EMBEDDING_LENGTH,
LLM_KV_POSNET_BLOCK_COUNT,
LLM_KV_CONVNEXT_EMBEDDING_LENGTH,
LLM_KV_CONVNEXT_BLOCK_COUNT,
LLM_KV_CLASSIFIER_OUTPUT_LABELS,
LLM_KV_CLASSIFIER_POOLING_TYPE,
LLM_KV_TARGET_LAYERS,
LLM_KV_TARGET_HIDDEN_SIZE,
LLM_KV_DFLASH_BLOCK_SIZE,
LLM_KV_DFLASH_CONV_KERNEL_SIZE,
LLM_KV_DFLASH_CONV_GROUP_SIZE,
LLM_KV_DFLASH_SELECTOR_RANK,
LLM_KV_DFLASH_SELECTOR_TOP_K,
LLM_KV_NORM_BEFORE_RESIDUAL,
LLM_KV_NORM_BEFORE_FC,
LLM_KV_SHORTCONV_L_CACHE,
LLM_KV_XIELU_ALPHA_N,
LLM_KV_XIELU_ALPHA_P,
LLM_KV_XIELU_BETA,
LLM_KV_XIELU_EPS,
// deprecated:
LLM_KV_TOKENIZER_PREFIX_ID,
LLM_KV_TOKENIZER_SUFFIX_ID,
LLM_KV_TOKENIZER_MIDDLE_ID,
// sentence-transformers dense layers in and out features
LLM_KV_DENSE_2_FEAT_IN,
LLM_KV_DENSE_2_FEAT_OUT,
LLM_KV_DENSE_3_FEAT_IN,
LLM_KV_DENSE_3_FEAT_OUT,
};
enum llm_tensor {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_TOKEN_EMBD_NORM,
LLM_TENSOR_TOKEN_TYPES,
LLM_TENSOR_POS_EMBD,
LLM_TENSOR_DENSE_2_OUT,
LLM_TENSOR_DENSE_3_OUT,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT_NORM_LFM2, // fix for wrong tensor name
LLM_TENSOR_ROPE_FREQS,
LLM_TENSOR_ROPE_FACTORS_LONG,
LLM_TENSOR_ROPE_FACTORS_SHORT,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_QKV,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_NORM_2,
LLM_TENSOR_ATTN_OUT_NORM,
LLM_TENSOR_ATTN_POST_NORM,
LLM_TENSOR_ATTN_ROT_EMBD,
LLM_TENSOR_ATTN_SINKS,
LLM_TENSOR_ATTN_GATE,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_GATE_INP_SHEXP,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_POST_NORM,
LLM_TENSOR_FFN_POST_NORM_1,
LLM_TENSOR_FFN_POST_NORM_2,
LLM_TENSOR_FFN_PRE_NORM_2,
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_UP,
LLM_TENSOR_FFN_ACT,
LLM_TENSOR_FFN_DOWN_EXP, // split experts for backward compatibility
LLM_TENSOR_FFN_GATE_EXP,
LLM_TENSOR_FFN_UP_EXP,
LLM_TENSOR_FFN_NORM_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS, // merged experts
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_GATE_UP_EXPS,
LLM_TENSOR_FFN_DOWN_SHEXP,
LLM_TENSOR_FFN_GATE_SHEXP,
LLM_TENSOR_FFN_UP_SHEXP,
LLM_TENSOR_FFN_DOWN_CHEXPS,
LLM_TENSOR_FFN_GATE_CHEXPS,
LLM_TENSOR_FFN_UP_CHEXPS,
LLM_TENSOR_FFN_EXP_PROBS_B,
LLM_TENSOR_FFN_EXP_PROBS_B_VL,
LLM_TENSOR_FFN_LATENT_DOWN,
LLM_TENSOR_FFN_LATENT_UP,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_LAYER_OUT_NORM,
LLM_TENSOR_LAYER_OUT_SCALE,
LLM_TENSOR_POST_ATTN_NORM,
LLM_TENSOR_POST_MLP_NORM,
LLM_TENSOR_PER_LAYER_TOKEN_EMBD, // gemma3n
LLM_TENSOR_PER_LAYER_MODEL_PROJ, // gemma3n
LLM_TENSOR_PER_LAYER_INP_GATE, // gemma3n
LLM_TENSOR_PER_LAYER_PROJ, // gemma3n
LLM_TENSOR_PER_LAYER_PROJ_NORM, // gemma3n
LLM_TENSOR_PER_LAYER_POST_NORM, // gemma3n
LLM_TENSOR_ALTUP_PROJ, // gemma3n
LLM_TENSOR_ALTUP_UNEMBD_PROJ, // gemma3n
LLM_TENSOR_ALTUP_CORRECT_COEF, // gemma3n
LLM_TENSOR_ALTUP_CORRECT_SCALE, // gemma3n
LLM_TENSOR_ALTUP_PREDICT_COEF, // gemma3n
LLM_TENSOR_ALTUP_ROUTER, // gemma3n
LLM_TENSOR_ALTUP_ROUTER_NORM, // gemma3n
LLM_TENSOR_LAUREL_L, // gemma3n
LLM_TENSOR_LAUREL_R, // gemma3n
LLM_TENSOR_LAUREL_POST_NORM, // gemma3n
LLM_TENSOR_SSM_IN,
LLM_TENSOR_SSM_CONV1D,
LLM_TENSOR_SSM_X,
LLM_TENSOR_SSM_DT,
LLM_TENSOR_SSM_DT_NORM,
LLM_TENSOR_SSM_A,
LLM_TENSOR_SSM_A_NOSCAN, // qwen3next special case with MUL instead of SSM_SCAN
LLM_TENSOR_SSM_B_NORM,
LLM_TENSOR_SSM_C_NORM,
LLM_TENSOR_SSM_D,
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
LLM_TENSOR_SSM_BETA_ALPHA, // qwen3next
LLM_TENSOR_SSM_ALPHA, // qwen3.5
// Kimi Linear KDA (using SSM_ prefix for consistency)
LLM_TENSOR_SSM_CONV1D_Q, // kimi: Q conv1d weight
LLM_TENSOR_SSM_CONV1D_K, // kimi: K conv1d weight
LLM_TENSOR_SSM_CONV1D_V, // kimi: V conv1d weight
LLM_TENSOR_SSM_F_A, // kimi: forget gate projection A
LLM_TENSOR_SSM_F_B, // kimi: forget gate projection B
LLM_TENSOR_SSM_BETA, // kimi: beta mixing coefficient and qwen3.5
LLM_TENSOR_SSM_G_A, // kimi: output gate projection A
LLM_TENSOR_SSM_G_B, // kimi: output gate projection B
LLM_TENSOR_SSM_G, // kimi-k3: full-rank KDA gate
LLM_TENSOR_ATTN_RES_SCORE, // kimi-k3: fused res_norm*res_proj (pre-attn)
LLM_TENSOR_FFN_RES_SCORE, // kimi-k3: fused res_norm*res_proj (pre-ffn)
LLM_TENSOR_OUTPUT_RES_SCORE, // kimi-k3: fused res_norm*res_proj (final)
LLM_TENSOR_FFN_ROUTED_DOWN, // kimi-k3: latent MoE down
LLM_TENSOR_FFN_ROUTED_UP, // kimi-k3: latent MoE up
LLM_TENSOR_FFN_ROUTED_NORM, // kimi-k3: latent MoE norm
LLM_TENSOR_TIME_MIX_W0,
LLM_TENSOR_TIME_MIX_W1,
LLM_TENSOR_TIME_MIX_W2,
LLM_TENSOR_TIME_MIX_A0,
LLM_TENSOR_TIME_MIX_A1,
LLM_TENSOR_TIME_MIX_A2,
LLM_TENSOR_TIME_MIX_V0,
LLM_TENSOR_TIME_MIX_V1,
LLM_TENSOR_TIME_MIX_V2,
LLM_TENSOR_TIME_MIX_G1,
LLM_TENSOR_TIME_MIX_G2,
LLM_TENSOR_TIME_MIX_K_K,
LLM_TENSOR_TIME_MIX_K_A,
LLM_TENSOR_TIME_MIX_R_K,
LLM_TENSOR_TIME_MIX_LERP_X,
LLM_TENSOR_TIME_MIX_LERP_W,
LLM_TENSOR_TIME_MIX_LERP_K,
LLM_TENSOR_TIME_MIX_LERP_V,
LLM_TENSOR_TIME_MIX_LERP_R,
LLM_TENSOR_TIME_MIX_LERP_G,
LLM_TENSOR_TIME_MIX_LERP_FUSED,
LLM_TENSOR_TIME_MIX_FIRST,
LLM_TENSOR_TIME_MIX_DECAY,
LLM_TENSOR_TIME_MIX_DECAY_W1,
LLM_TENSOR_TIME_MIX_DECAY_W2,
LLM_TENSOR_TIME_MIX_KEY,
LLM_TENSOR_TIME_MIX_VALUE,
LLM_TENSOR_TIME_MIX_RECEPTANCE,
LLM_TENSOR_TIME_MIX_GATE,
LLM_TENSOR_TIME_MIX_LN,
LLM_TENSOR_TIME_MIX_OUTPUT,
LLM_TENSOR_CHANNEL_MIX_LERP_K,
LLM_TENSOR_CHANNEL_MIX_LERP_R,
LLM_TENSOR_CHANNEL_MIX_KEY,
LLM_TENSOR_CHANNEL_MIX_RECEPTANCE,
LLM_TENSOR_CHANNEL_MIX_VALUE,
LLM_TENSOR_ATTN_Q_A,
LLM_TENSOR_ATTN_Q_B,
LLM_TENSOR_ATTN_KV_A_MQA,
LLM_TENSOR_ATTN_KV_B,
LLM_TENSOR_ATTN_KV,
LLM_TENSOR_ATTN_KV_NORM,
LLM_TENSOR_ATTN_OUT_A,
LLM_TENSOR_ATTN_OUT_B,
LLM_TENSOR_ATTN_K_B,
LLM_TENSOR_ATTN_V_B,
LLM_TENSOR_ATTN_Q_A_NORM,
LLM_TENSOR_ATTN_KV_A_NORM,
LLM_TENSOR_HC_HEAD_FN,
LLM_TENSOR_HC_HEAD_BASE,
LLM_TENSOR_HC_HEAD_SCALE,
LLM_TENSOR_HC_HEAD_NORM, // qwen4exp
LLM_TENSOR_HC_HEAD_DOWN, // qwen4exp
LLM_TENSOR_HC_HEAD_UP, // qwen4exp
LLM_TENSOR_HC_ATTN_FN,
LLM_TENSOR_HC_ATTN_BASE,
LLM_TENSOR_HC_ATTN_SCALE,
LLM_TENSOR_HC_FFN_FN,
LLM_TENSOR_HC_FFN_BASE,
LLM_TENSOR_HC_FFN_SCALE,
LLM_TENSOR_HC_ATTN_NORM, // qwen4exp
LLM_TENSOR_HC_ATTN_DOWN, // qwen4exp
LLM_TENSOR_HC_ATTN_UP, // qwen4exp
LLM_TENSOR_HC_ATTN_INJECT, // qwen4exp
LLM_TENSOR_HC_FFN_NORM, // qwen4exp
LLM_TENSOR_HC_FFN_DOWN, // qwen4exp
LLM_TENSOR_HC_FFN_UP, // qwen4exp
LLM_TENSOR_HC_FFN_INJECT, // qwen4exp
LLM_TENSOR_PLE_KEY, // qwen4exp
LLM_TENSOR_PLE_VALUE, // qwen4exp
LLM_TENSOR_PLE_NORM_KEY, // qwen4exp
LLM_TENSOR_PLE_NORM_QUERY, // qwen4exp
LLM_TENSOR_PLE_NORM_CONV, // qwen4exp
LLM_TENSOR_PLE_CONV1D, // qwen4exp
LLM_TENSOR_ATTN_COMPRESSOR_WKV,
LLM_TENSOR_ATTN_COMPRESSOR_WGATE,
LLM_TENSOR_ATTN_COMPRESSOR_APE,
LLM_TENSOR_ATTN_COMPRESSOR_NORM,
LLM_TENSOR_ATTN_SUB_NORM,
LLM_TENSOR_FFN_SUB_NORM,
LLM_TENSOR_DEC_ATTN_NORM,
LLM_TENSOR_DEC_ATTN_Q,
LLM_TENSOR_DEC_ATTN_K,
LLM_TENSOR_DEC_ATTN_V,
LLM_TENSOR_DEC_ATTN_OUT,
LLM_TENSOR_DEC_ATTN_REL_B,
LLM_TENSOR_DEC_CROSS_ATTN_NORM,
LLM_TENSOR_DEC_CROSS_ATTN_Q,
LLM_TENSOR_DEC_CROSS_ATTN_K,
LLM_TENSOR_DEC_CROSS_ATTN_V,
LLM_TENSOR_DEC_CROSS_ATTN_OUT,
LLM_TENSOR_DEC_CROSS_ATTN_REL_B,
LLM_TENSOR_DEC_FFN_NORM,
LLM_TENSOR_DEC_FFN_GATE,
LLM_TENSOR_DEC_FFN_DOWN,
LLM_TENSOR_DEC_FFN_UP,
LLM_TENSOR_DEC_OUTPUT_NORM,
LLM_TENSOR_ENC_ATTN_NORM,
LLM_TENSOR_ENC_ATTN_Q,
LLM_TENSOR_ENC_ATTN_K,
LLM_TENSOR_ENC_ATTN_V,
LLM_TENSOR_ENC_ATTN_OUT,
LLM_TENSOR_ENC_ATTN_REL_B,
LLM_TENSOR_ENC_FFN_NORM,
LLM_TENSOR_ENC_FFN_GATE,
LLM_TENSOR_ENC_FFN_DOWN,
LLM_TENSOR_ENC_FFN_UP,
LLM_TENSOR_ENC_OUTPUT_NORM,
LLM_TENSOR_CLS,
LLM_TENSOR_CLS_OUT,
LLM_TENSOR_CLS_NORM,
LLM_TENSOR_CONV1D,
LLM_TENSOR_CONVNEXT_DW,
LLM_TENSOR_CONVNEXT_NORM,
LLM_TENSOR_CONVNEXT_PW1,
LLM_TENSOR_CONVNEXT_PW2,
LLM_TENSOR_CONVNEXT_GAMMA,
LLM_TENSOR_POS_NET_CONV1,
LLM_TENSOR_POS_NET_CONV2,
LLM_TENSOR_POS_NET_NORM,
LLM_TENSOR_POS_NET_NORM1,
LLM_TENSOR_POS_NET_NORM2,
LLM_TENSOR_POS_NET_ATTN_NORM,
LLM_TENSOR_POS_NET_ATTN_Q,
LLM_TENSOR_POS_NET_ATTN_K,
LLM_TENSOR_POS_NET_ATTN_V,
LLM_TENSOR_POS_NET_ATTN_OUT,
LLM_TENSOR_SHORTCONV_CONV,
LLM_TENSOR_SHORTCONV_INPROJ,
LLM_TENSOR_SHORTCONV_OUTPROJ,
LLM_TENSOR_VISEXP_ATTN_QKV,
LLM_TENSOR_VISEXP_ATTN_OUT,
LLM_TENSOR_VISEXP_FFN_GATE,
LLM_TENSOR_VISEXP_FFN_DOWN,
LLM_TENSOR_VISEXP_FFN_UP,
LLM_TENSOR_INDEXER_K_NORM,
LLM_TENSOR_INDEXER_PROJ,
LLM_TENSOR_INDEXER_ATTN_K,
LLM_TENSOR_INDEXER_ATTN_Q_B,
LLM_TENSOR_INDEXER_Q_PROJ,
LLM_TENSOR_INDEXER_K_PROJ,
LLM_TENSOR_INDEXER_Q_NORM,
LLM_TENSOR_INDEXER_COMPRESSOR_WKV,
LLM_TENSOR_INDEXER_COMPRESSOR_WGATE,
LLM_TENSOR_INDEXER_COMPRESSOR_APE,
LLM_TENSOR_INDEXER_COMPRESSOR_NORM,
LLM_TENSOR_INDEXER_KPOOL_GATE, // glm5-next: k-pool gate scores
LLM_TENSOR_INDEXER_KPOOL_APE, // glm5-next: k-pool position bias
LLM_TENSOR_FFN_GATE_TID2EID,
LLM_TENSOR_NEXTN_PROJ_PRE,
LLM_TENSOR_NEXTN_PROJ_POST,
LLM_TENSOR_NEXTN_EH_PROJ,
LLM_TENSOR_NEXTN_EMBED_TOKENS,
LLM_TENSOR_NEXTN_ENORM,
LLM_TENSOR_NEXTN_HNORM,
LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
LLM_TENSOR_MASKED_EMBD_CENTROIDS,
LLM_TENSOR_MASKED_EMBD_ORDERING,
LLM_TENSOR_HRM_Z_L_INIT,
LLM_TENSOR_FC,
LLM_TENSOR_D2T,
LLM_TENSOR_DSPARK_MARKOV_W1,
LLM_TENSOR_DSPARK_MARKOV_W2,
LLM_TENSOR_DSPARK_CONF_PROJ,
LLM_TENSOR_DFLASH_ATTN_CONV_BASE,
LLM_TENSOR_DFLASH_ATTN_CONV_PROJ,
LLM_TENSOR_DFLASH_FFN_CONV_BASE,
LLM_TENSOR_DFLASH_FFN_CONV_PROJ,
LLM_TENSOR_DFLASH_SELECTOR_PREV,
LLM_TENSOR_DFLASH_SELECTOR_NEXT,
LLM_TENSOR_DFLASH_SELECTOR_HIDDEN,
};
enum llm_tensor_layer {
LLM_TENSOR_LAYER_INPUT,
LLM_TENSOR_LAYER_REPEATING,
LLM_TENSOR_LAYER_OUTPUT,
};
struct LLM_KV {
LLM_KV(llm_arch arch, const char * suffix = nullptr);
llm_arch arch;
const char * suffix;
std::string operator()(llm_kv kv) const;
};
// helper to handle gguf constants
// usage:
//
// const auto tn = LLM_TN(LLM_ARCH_LLAMA);
//
// std::string name = tn(LLM_TENSOR_OUTPUT); -> "output"
// std::string name = tn(LLM_TENSOR_TOKEN_EMBD, "bias"); -> "token_embd.bias"
// std::string name = tn(LLM_TENSOR_ATTN_NORM, "weight", 3); -> "blk.3.attn_norm.weight"
//
struct LLM_TN_IMPL {
const llm_arch arch;
const llm_tensor tensor;
const char * const suffix;
const int bid;
const int xid;
LLM_TN_IMPL(llm_arch arch, llm_tensor tensor, const char * suffix, int bid, int xid);
std::string str() const;
operator std::string() const {
return str();
}
friend bool operator==(const std::string & str, const LLM_TN_IMPL & tn) {
return str == tn.str();
}
friend bool operator!=(const std::string & str, const LLM_TN_IMPL & tn) {
return str != tn.str();
}
};
struct LLM_TN {
LLM_TN(llm_arch arch) : arch(arch) {}
llm_arch arch;
LLM_TN_IMPL operator()(llm_tensor tensor, const char * suffix, int bid = -1, int xid = -1) const {
return LLM_TN_IMPL(arch, tensor, suffix, bid, xid);
}
LLM_TN_IMPL operator()(llm_tensor tensor, int bid = -1, int xid = -1) const {
return LLM_TN_IMPL(arch, tensor, nullptr, bid, xid);
}
};
struct llm_tensor_info {
llm_tensor_layer layer;
ggml_op op;
};
std::vector<llm_arch> llm_arch_all();
const char * llm_arch_name(llm_arch arch);
llm_arch llm_arch_from_string(const std::string & name);
const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor);
bool llm_arch_is_recurrent (const llm_arch & arch);
bool llm_arch_is_hybrid (const llm_arch & arch);
bool llm_arch_is_diffusion (const llm_arch & arch);
bool llm_arch_supports_sm_tensor(const llm_arch & arch);
bool llm_arch_supports_rs_rollback(const llm_arch & arch);
Reference in New Issue View Git Blame Copy Permalink