[Kernel] Gemma-4 FA4 FP8 Kernel (#48666)

Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
Signed-off-by: Jhao-Ting Chen <jhaotingc@nvidia.com>
Co-authored-by: Qixiang Lin <qixiangl@nvidia.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
Jhao-Ting Chen
2026-08-13 17:42:00 -07:00
committed by GitHub
co-authored by Qixiang Lin Claude Opus 4.8 Yongye Zhu
parent 38f097fab8
commit 827a2af806
8 changed files with 152 additions and 29 deletions
+3 -5
View File
@@ -612,7 +612,6 @@ class Platform:
For hybrid models, also aligns block_size with mamba page sizes.
"""
from vllm.config.cache import CacheConfig
from vllm.config.vllm import set_current_vllm_config
cache_config = vllm_config.cache_config
model_config = vllm_config.model_config
@@ -627,10 +626,9 @@ class Platform:
# Phase 1: Pick block size from backend (skip if user set --block-size)
if not cache_config.user_specified_block_size:
with set_current_vllm_config(vllm_config):
preferred = backend_cls.get_preferred_block_size(
CacheConfig.DEFAULT_BLOCK_SIZE
)
preferred = backend_cls.get_preferred_block_size_for_config(
CacheConfig.DEFAULT_BLOCK_SIZE, vllm_config
)
if preferred != CacheConfig.DEFAULT_BLOCK_SIZE:
logger.info(
"Setting kv cache block size to %d for %s backend.",