[Experimental] Breakable CUDA graph (#42304)

Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
This commit is contained in:
Jiangyun Zhu
2026-05-16 22:04:12 +08:00
committed by GitHub
parent 4db300e95f
commit 8a56da3845
11 changed files with 845 additions and 5 deletions
@@ -200,6 +200,7 @@ from tqdm import tqdm
import vllm.envs as envs
from vllm import _custom_ops as ops
from vllm._aiter_ops import rocm_aiter_ops
from vllm.compilation.breakable_cudagraph import eager_break_during_capture
from vllm.config import (
CacheConfig,
ModelConfig,
@@ -1036,6 +1037,7 @@ direct_register_custom_op(
)
@eager_break_during_capture
@maybe_transfer_kv_layer
def unified_mla_attention_with_output(
q: torch.Tensor,
@@ -14,6 +14,7 @@ import torch.nn.functional as F
from transformers import DeepseekV2Config, DeepseekV3Config
import vllm.envs as envs
from vllm.compilation.breakable_cudagraph import eager_break_during_capture
from vllm.model_executor.layers.linear import (
ReplicatedLinear,
)
@@ -553,6 +554,7 @@ class DeepseekV4MultiHeadLatentAttentionWrapper(PluggableLayer):
)
@eager_break_during_capture
def deepseek_v4_attention(
hidden_states: torch.Tensor,
positions: torch.Tensor,
@@ -6,6 +6,7 @@ import torch
import vllm.envs as envs
from vllm._aiter_ops import rocm_aiter_ops
from vllm.compilation.breakable_cudagraph import eager_break_during_capture
from vllm.forward_context import get_forward_context
from vllm.logger import init_logger
from vllm.model_executor.custom_op import CustomOp
@@ -81,6 +82,7 @@ def kv_cache_as_quant_view(
return kv_cache.unsqueeze(-2)
@eager_break_during_capture
def sparse_attn_indexer(
hidden_states: torch.Tensor,
k_cache_prefix: LayerNameType,