[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
@@ -10,6 +10,7 @@ from vllm.model_executor.layers.quantization.base_config import (
)
from vllm.platforms import current_platform
from vllm.utils.torch_utils import is_quantized_kv_cache
from vllm.v1.kv_cache_interface import kv_cache_uses_per_token_head_scales
logger = init_logger(__name__)
@@ -53,6 +54,20 @@ class BaseKVCacheMethod(QuantizeMethodBase):
assert not hasattr(layer, "prob_scale")
return
# Per-token-head quantized KV cache: scales are computed dynamically
# per (token, head) in the kernel at cache-write time. Checkpoint
# scales are never used regardless of calculate_kv_scales.
if kv_cache_uses_per_token_head_scales(layer.kv_cache_dtype):
layer._k_scale.copy_(1.0)
layer._v_scale.copy_(1.0)
layer._k_scale_float = 1.0
layer._v_scale_float = 1.0
del layer.k_scale
del layer.v_scale
del layer.q_scale
del layer.prob_scale
return
# If the kv-cache is not quantized, we enforce the k/v_scale to be 1.0
# regardless whether the kv-scale is available in the checkpoint.
# No need to process kv scales after loading if we are going to