Remove KV cache scale boilerplate from model weight loading methods (#43167)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-06-05 05:19:04 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 7fe7800fa4
commit 62215e72c6
56 changed files with 87 additions and 730 deletions
+11 -18
View File
@@ -207,25 +207,18 @@ class Fp8Config(QuantizationConfig):
return Fp8KVCacheMethod(self)
return None
def get_cache_scale(self, name: str) -> str | None:
"""
Check whether the param name matches the format for k/v cache scales
in compressed-tensors. If this is the case, return its equivalent
param name expected by vLLM
def get_cache_scale_mapper(self) -> "WeightsMapper":
"""Map compressed-tensors KV-cache scale names to vLLM names."""
from vllm.model_executor.models.utils import WeightsMapper
:param name: param name
:return: matching param name for KV cache scale in vLLM
"""
if name.endswith(".output_scale") and ".k_proj" in name:
return name.replace(".k_proj.output_scale", ".attn.k_scale")
if name.endswith(".output_scale") and ".v_proj" in name:
return name.replace(".v_proj.output_scale", ".attn.v_scale")
if name.endswith(".output_scale") and ".q_proj" in name:
return name.replace(".q_proj.output_scale", ".attn.q_scale")
if name.endswith("self_attn.prob_output_scale"):
return name.replace(".prob_output_scale", ".attn.prob_scale")
# If no matches, return None
return None
return WeightsMapper(
orig_to_new_suffix={
".k_proj.output_scale": ".attn.k_scale",
".v_proj.output_scale": ".attn.v_scale",
".q_proj.output_scale": ".attn.q_scale",
".self_attn.prob_output_scale": ".self_attn.attn.prob_scale",
}
)
class CopyNumelCounter(TorchDispatchMode):