[Quantization/NVFP4] Speed up TRTLLM NVFP4 MOE weight loading and fix K/V scale loading for MLA Attn (#25968)

Signed-off-by: Pavani Majety <pmajety@nvidia.com>
This commit is contained in:
Pavani Majety
2025-10-03 19:35:06 +00:00
committed by GitHub
parent cd9e5b8340
commit a26917332f
3 changed files with 77 additions and 55 deletions
@@ -1003,12 +1003,18 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> Optional[str]:
return None
return remapped_name
if any("mla_attn" in key for key in params_dict):
attn_str = "mla_attn.mla_attn"
logger.debug_once(f"Found mla_attn with k_scale and v_scale in "
f"the checkpoint, using {attn_str} as attn_str")
else:
attn_str = "attn"
# Define scale name mapping patterns in order of precedence
scale_mapping_patterns = [
# ModelOpt format: .self_attn.{k,v}_proj.{k,v}_scale ->
# .self_attn.attn.{k,v}_scale
(r"\.self_attn\.([kv])_proj\.([kv])_scale$",
r".self_attn.attn.\2_scale"),
rf".self_attn.{attn_str}.\2_scale"),
# QKV proj format: .self_attn.qkv_proj.{k,v}_scale ->
# .self_attn.attn.{k,v}_scale
(r"\.self_attn\.qkv_proj\.([kv])_scale$", r".self_attn.attn.\1_scale"),