[BreakableCG] Add eager break for Mamba-family ops (#57881)

Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
liuzhenwei
2026-10-10 14:06:57 +08:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent a205fb3177
commit ff440c7805
5 changed files with 10 additions and 0 deletions
@@ -4,6 +4,7 @@ import torch
from einops import rearrange
from torch import nn
from vllm.compilation.breakable_cudagraph import eager_break_during_capture
from vllm.config import (
VllmConfig,
get_current_vllm_config,
@@ -524,6 +525,7 @@ def _make_fused_conv1d_weight_loader(dims, tp_size, tp_rank):
return weight_loader
@eager_break_during_capture
def olmo_hybrid_gdn_full_forward(
hidden_states: torch.Tensor,
output: torch.Tensor,
@@ -8,6 +8,7 @@ import torch
import torch.nn.functional as F
from einops import rearrange
from vllm.compilation.breakable_cudagraph import eager_break_during_capture
from vllm.config import get_current_vllm_config
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.model_executor.custom_op import PluggableLayer
@@ -325,6 +326,7 @@ class MiniMaxText01LinearAttention(LinearAttention):
output[:num_actual_tokens], _ = self.out_proj(hidden)
@eager_break_during_capture
def linear_attention(
hidden_states: torch.Tensor,
output: torch.Tensor,
@@ -7,6 +7,7 @@ import torch
from torch import nn
from torch.nn.parameter import Parameter
from vllm.compilation.breakable_cudagraph import eager_break_during_capture
from vllm.config import CacheConfig, ModelConfig, get_current_vllm_config
from vllm.distributed.parallel_state import (
get_tensor_model_parallel_rank,
@@ -465,6 +466,7 @@ def split_batch_to_prefill_and_decode(
)
@eager_break_during_capture
def mamba_mixer(
hidden_states: torch.Tensor,
output: torch.Tensor,
@@ -8,6 +8,7 @@ from dataclasses import replace
import torch
from torch import nn
from vllm.compilation.breakable_cudagraph import eager_break_during_capture
from vllm.config import CacheConfig, ModelConfig, VllmConfig, get_current_vllm_config
from vllm.config.mamba import MambaBackendEnum
from vllm.distributed import (
@@ -1214,6 +1215,7 @@ def share_replayssm_ring_trackers(
replayssm_mixers[last_layer_name]._updates_replayssm_trackers = True
@eager_break_during_capture
def mamba_mixer2(
projected_states: torch.Tensor,
output: torch.Tensor,
@@ -4,6 +4,7 @@
import torch
from vllm.compilation.breakable_cudagraph import eager_break_during_capture
from vllm.config import CacheConfig, ModelConfig, get_current_vllm_config
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.forward_context import ForwardContext, get_forward_context
@@ -360,6 +361,7 @@ class ShortConv(MambaBase, PluggableLayer):
return MambaAttentionBackendEnum.SHORT_CONV
@eager_break_during_capture
def short_conv(
hidden_states: torch.Tensor,
output: torch.Tensor,