minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[AMD] Fix Quark load of MiniMax-M3 MXFP4 index_qkv_proj (#37254)

Browse Source
This commit is contained in:
Tuan Nguyen Gia
2026-09-12 00:50:41 -07:00
committed by GitHub
parent bf3305b65e
commit 7c195b9151
3 changed files with 54 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/minimax_m3.py
+2 -1
View File
@@ -1559,7 +1559,8 @@ class MiniMaxM3SparseForCausalLM(nn.Module):
)
packed_modules_mapping = {
"qkv_proj": ["q_proj", "k_proj", "v_proj"],
"index_qkv_proj": ["index_q_proj", "index_k_proj", "index_v_proj"],
# no index_v_proj in the M3 checkpoint
"index_qkv_proj": ["index_q_proj", "index_k_proj"],
"gate_up_proj": ["gate_proj", "up_proj"],
}
python/sglang/srt/models/minimax_m3_vl.py
+2 -1
View File
@@ -63,7 +63,8 @@ class MiniMaxM3SparseForConditionalGeneration(nn.Module):
)
packed_modules_mapping = {
"qkv_proj": ["q_proj", "k_proj", "v_proj"],
"index_qkv_proj": ["index_q_proj", "index_k_proj", "index_v_proj"],
# no index_v_proj in the M3 checkpoint
"index_qkv_proj": ["index_q_proj", "index_k_proj"],
"gate_up_proj": ["gate_proj", "up_proj"],
}