Revert "[Kernel] Gemma-4 FA4 FP8 Kernel" (#52987)

This commit is contained in:
Roger Wang
2026-08-19 10:48:46 -07:00
committed by GitHub
parent f76d71d7ce
commit 755492e37d
8 changed files with 29 additions and 152 deletions
@@ -8,10 +8,7 @@ import torch.nn as nn
import vllm.envs as envs
from vllm.compilation.breakable_cudagraph import eager_break_during_capture
from vllm.config import (
CacheConfig,
get_current_vllm_config,
)
from vllm.config import CacheConfig, get_current_vllm_config
from vllm.config.vllm import VllmConfig
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.logger import init_logger
@@ -97,7 +94,6 @@ def should_load_quant_weights(quant_method: QuantizeMethodBase | None) -> bool:
def _largest_kernel_block_within(
attn_backend: "type[AttentionBackend]",
vllm_config: VllmConfig,
per_token_bytes: int,
page_budget: int | None,
fallback: int,
@@ -112,7 +108,7 @@ def _largest_kernel_block_within(
"""
from vllm.v1.attention.backend import MultipleOf
sizes = attn_backend.get_supported_kernel_block_sizes_for_config(vllm_config)
sizes = attn_backend.get_supported_kernel_block_sizes()
candidates = [s for s in sizes if isinstance(s, int)]
if not candidates:
candidates = [s.base for s in sizes if isinstance(s, MultipleOf)]
@@ -636,11 +632,7 @@ class Attention(nn.Module, AttentionLayerBase):
)
).real_page_size_bytes
sw_block_size = _largest_kernel_block_within(
self.attn_backend,
vllm_config,
sw_per_token,
shared_page,
block_size,
self.attn_backend, sw_per_token, shared_page, block_size
)
return SlidingWindowSpec(
block_size=sw_block_size,