[Feature] KV cache per-token-head INT8/FP8 quantization (#38378)

Signed-off-by: JartX <sagformas@epdcenter.es>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: yangyang4991 <yangyang4991@gmail.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: Isotr0py <2037008807@qq.com>
This commit is contained in:
JartX
2026-04-02 08:13:26 -04:00
committed by GitHub
co-authored by yangyang4991 Wentao Ye Isotr0py
parent 4eefbf9609
commit 2ce3d0ce36
16 changed files with 1308 additions and 66 deletions
+17 -4
View File
@@ -8,7 +8,10 @@ from pydantic import Field, SkipValidation, field_validator, model_validator
from vllm.config.utils import config
from vllm.logger import init_logger
from vllm.utils.torch_utils import is_quantized_kv_cache
from vllm.utils.torch_utils import (
is_quantized_kv_cache,
kv_cache_uses_per_token_head_scales,
)
logger = init_logger(__name__)
@@ -21,6 +24,8 @@ CacheDType = Literal[
"fp8_e5m2",
"fp8_inc",
"fp8_ds_mla",
"int8_per_token_head",
"fp8_per_token_head",
]
MambaDType = Literal["auto", "float32", "float16"]
MambaCacheMode = Literal["all", "align", "none"]
@@ -237,12 +242,20 @@ class CacheConfig:
@field_validator("cache_dtype", mode="after")
@classmethod
def _validate_cache_dtype(cls, cache_dtype: CacheDType) -> CacheDType:
if is_quantized_kv_cache(cache_dtype):
if kv_cache_uses_per_token_head_scales(cache_dtype):
logger.info(
"Using fp8 data type to store kv cache. It reduces the GPU "
"Using %s data type to store kv cache. It reduces the GPU "
"memory footprint and boosts the performance. "
"Dynamic per-token-head scales will be computed at runtime.",
str(cache_dtype),
)
elif is_quantized_kv_cache(cache_dtype):
logger.info(
"Using %s data type to store kv cache. It reduces the GPU "
"memory footprint and boosts the performance. "
"Meanwhile, it may cause accuracy drop without a proper "
"scaling factor."
"scaling factor",
str(cache_dtype),
)
return cache_dtype