vrr/kvcache-ai-ktransformers
2
0
Fork 0
mirror of https://github.com/kvcache-ai/ktransformers.git synced 2026-05-05 07:11:39 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

update: Qwen3 MoE model adaptation for NPU (framework) (#1706)

Browse source
This commit is contained in:
Shaoxu Cheng 2025-12-11 17:07:57 +08:00 • committed by GitHub
parent 53f6a6d6e1
commit adcfa9080f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
10 changed files with 867 additions and 174 deletions
Download patch file Download diff file Expand all files Collapse all files

27
archive/ktransformers/util/custom_gguf.py
View file

@ -666,6 +666,33 @@ def translate_name_to_gguf(name):
name = translate_name_to_gguf_mixtral(name)
if ".ffn_gate_exp." in name:
name = name.replace(".ffn_gate_exp.", ".ffn_gate_exps.")
if ".ffn_up_exp." in name:
name = name.replace(".ffn_up_exp.", ".ffn_up_exps.")
if ".ffn_down_exp." in name:
name = name.replace(".ffn_down_exp.", ".ffn_down_exps.")
m = re.match(r"model\.layers\.(\d+)\.mlp\.experts\.(\d+)\.(gate_proj|up_proj|down_proj)", name)
if m:
layer, expert, proj = m.groups()
if proj == "gate_proj":
return f"blk.{layer}.{expert}.ffn_gate_exps"
elif proj == "up_proj":
return f"blk.{layer}.{expert}.ffn_up_exps"
else:
return f"blk.{layer}.{expert}.ffn_down_exps"
m = re.match(r"blk\.(\d+)\.mlp\.experts\.(\d+)\.(gate_proj|up_proj|down_proj)", name)
if m:
layer, expert, proj = m.groups()
if proj == "gate_proj":
return f"blk.{layer}.{expert}.ffn_gate_exps"
elif proj == "up_proj":
return f"blk.{layer}.{expert}.ffn_up_exps"
else:
return f"blk.{layer}.{expert}.ffn_down_exps"
name = name.replace("lm_head.", "output.")
name = name.replace("model.embed_tokens.", "token_embd.")
name = name.replace("model.norm.", "output_norm.")