upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-07-23 11:10:55 -05:00
Code Issues Packages Projects Releases Wiki Activity

llama : add Gemma4 MTP (#23398)

Browse Source
This commit is contained in:
Aman Gupta
2026-06-07 20:50:54 +08:00
committed by GitHub
parent 8a091c47ab
commit 04eb4c446d
31 changed files with 644 additions and 142 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
gguf-py/gguf/constants.py
View File

@@ -440,6 +440,7 @@ class MODEL_ARCH(IntEnum):
GEMMA3 = auto()
GEMMA3N = auto()
GEMMA4 = auto()
GEMMA4_ASSISTANT = auto()
GEMMA_EMBEDDING = auto()
STARCODER2 = auto()
RWKV6 = auto()
@@ -897,6 +898,8 @@ class MODEL_TENSOR(IntEnum):
A_PER_DIM_K_SCALE = auto() # gemma4
A_PER_DIM_SCALE = auto() # gemma4
# nextn/mtp
NEXTN_PROJ_PRE = auto()
NEXTN_PROJ_POST = auto()
NEXTN_EH_PROJ = auto()
NEXTN_EMBED_TOKENS = auto()
NEXTN_ENORM = auto()
@@ -986,6 +989,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.GEMMA3: "gemma3",
MODEL_ARCH.GEMMA3N: "gemma3n",
MODEL_ARCH.GEMMA4: "gemma4",
MODEL_ARCH.GEMMA4_ASSISTANT: "gemma4-assistant",
MODEL_ARCH.GEMMA_EMBEDDING: "gemma-embedding",
MODEL_ARCH.STARCODER2: "starcoder2",
MODEL_ARCH.RWKV6: "rwkv6",
@@ -1471,6 +1475,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.A_QF_FFN_DOWN: "a.proj_blk.{bid}.ffn_down",
MODEL_TENSOR.A_QF_FFN_NORM: "a.proj_blk.{bid}.ffn_norm",
# NextN/MTP
MODEL_TENSOR.NEXTN_PROJ_PRE: "nextn.pre_projection",
MODEL_TENSOR.NEXTN_PROJ_POST: "nextn.post_projection",
MODEL_TENSOR.NEXTN_EH_PROJ: "blk.{bid}.nextn.eh_proj",
MODEL_TENSOR.NEXTN_EMBED_TOKENS: "blk.{bid}.nextn.embed_tokens",
MODEL_TENSOR.NEXTN_ENORM: "blk.{bid}.nextn.enorm",
@@ -2577,6 +2583,24 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.PER_LAYER_PROJ_NORM,
MODEL_TENSOR.PER_LAYER_POST_NORM,
],
MODEL_ARCH.GEMMA4_ASSISTANT: [
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.NEXTN_PROJ_PRE,
MODEL_TENSOR.NEXTN_PROJ_POST,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_POST_NORM,
MODEL_TENSOR.FFN_PRE_NORM,
MODEL_TENSOR.FFN_POST_NORM,
MODEL_TENSOR.LAYER_OUT_SCALE,
],
MODEL_ARCH.GEMMA_EMBEDDING: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT,

8
gguf-py/gguf/tensor_mapping.py
View File

@@ -2367,6 +2367,14 @@ class TensorNameMap:
),
# NextN/MTP tensors
MODEL_TENSOR.NEXTN_PROJ_PRE: (
"pre_projection",
),
MODEL_TENSOR.NEXTN_PROJ_POST: (
"post_projection",
),
MODEL_TENSOR.NEXTN_EH_PROJ: (
"model.layers.{bid}.eh_proj",
),