ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 19:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Support for dots.llm1 models (#573)

Browse Source 
* Add llama.cpp changes for dots1 support

* Add python changes for dots1 support

* Fix to make it convert

* Remove V reshaping, remove BOS by default for dots1 and fix warmup to handle models without BOS

* Minor fix

* Remove commented lines
This commit is contained in:
saood06
2025-07-10 02:37:36 -05:00
committed by GitHub
parent 283753cabc
commit c53cb65251
4 changed files with 326 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
gguf-py/gguf/constants.py
View File

@@ -226,6 +226,7 @@ class MODEL_ARCH(IntEnum):
T5 = auto()
T5ENCODER = auto()
JAIS = auto()
DOTS1 = auto()
class MODEL_TENSOR(IntEnum):
@@ -362,6 +363,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.T5: "t5",
MODEL_ARCH.T5ENCODER: "t5encoder",
MODEL_ARCH.JAIS: "jais",
MODEL_ARCH.DOTS1: "dots1",
}
TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -1164,6 +1166,30 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.DOTS1: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_EXP_PROBS_B,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_SHEXP,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.FFN_UP_SHEXP,
],
# TODO
}

2
gguf-py/gguf/tensor_mapping.py
View File

@@ -257,7 +257,7 @@ class TensorNameMap:
),
MODEL_TENSOR.FFN_EXP_PROBS_B: (
"model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3
"model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1
),
# Feed-forward up