[CI] Fix various failures on main (#47197)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-07-01 10:35:34 +01:00
committed by GitHub
parent aa8bb5562e
commit e7d0fcbc09
5 changed files with 18 additions and 14 deletions
+8 -13
View File
@@ -37,23 +37,18 @@ from vllm.platforms import current_platform
from vllm.utils.network_utils import get_open_port from vllm.utils.network_utils import get_open_port
def _weight_transfer_ray_env_vars() -> dict[str, str]:
if not current_platform.is_rocm():
return {}
return {
"RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES": "1",
"RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES": "1",
"RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES": "1",
}
def _init_ray_for_weight_transfer() -> None: def _init_ray_for_weight_transfer() -> None:
if ray.is_initialized(): if ray.is_initialized():
return return
ray.init( ray.init(
ignore_reinit_error=True, ignore_reinit_error=True,
runtime_env={"env_vars": _weight_transfer_ray_env_vars()}, runtime_env={
"env_vars": {
"RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES": "1",
"RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES": "1",
"RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES": "1",
}
},
) )
@@ -66,7 +61,7 @@ def _get_ray_assigned_device() -> torch.device:
def _set_ray_assigned_device() -> torch.device: def _set_ray_assigned_device() -> torch.device:
device = _get_ray_assigned_device() device = _get_ray_assigned_device()
torch.accelerator.set_device_index(device) current_platform.set_device(device)
return device return device
@@ -403,6 +403,7 @@ def _run_ref_mamba_state_worker():
GPUModelRunner._sample = fake_sample_fn GPUModelRunner._sample = fake_sample_fn
engine = LLM( engine = LLM(
model=MODEL, model=MODEL,
load_format="dummy",
block_size=BLOCK_SIZE, block_size=BLOCK_SIZE,
hf_overrides={"num_hidden_layers": NUM_HIDDEN_LAYERS}, hf_overrides={"num_hidden_layers": NUM_HIDDEN_LAYERS},
seed=42, seed=42,
@@ -750,6 +751,7 @@ def test_mamba_prefix_cache_mrv1(monkeypatch: pytest.MonkeyPatch):
engine = LLM( engine = LLM(
model=MODEL, model=MODEL,
load_format="dummy",
enable_prefix_caching=True, enable_prefix_caching=True,
block_size=BLOCK_SIZE, block_size=BLOCK_SIZE,
mamba_cache_mode="align", mamba_cache_mode="align",
@@ -907,7 +907,7 @@ class RoutedExperts(PluggableLayer):
# Unified loading logic for fused and non-fused experts # Unified loading logic for fused and non-fused experts
loaded_experts = experts_shard.unbind() loaded_experts = experts_shard.unbind()
for expert_id, loaded_expert in enumerate(loaded_experts, start=start): for expert_id, loaded_expert in enumerate(loaded_experts, start=start):
success = self.weight_loader( success = param.weight_loader(
param=param, param=param,
loaded_weight=loaded_expert, loaded_weight=loaded_expert,
weight_name=weight_name, weight_name=weight_name,
@@ -1445,6 +1445,8 @@ class DeepseekV2Model(nn.Module):
hidden_states, residual = combined_states.split( hidden_states, residual = combined_states.split(
[self.hidden_size, self.hidden_size], dim=-1 [self.hidden_size, self.hidden_size], dim=-1
) )
# fused_add_rms_norm requires a contiguous residual
residual = residual.contiguous()
if idx in self.aux_hidden_state_layers: if idx in self.aux_hidden_state_layers:
aux_hidden_state = hidden_states + residual aux_hidden_state = hidden_states + residual
if aux_hidden_state.shape[0] != positions.shape[0]: if aux_hidden_state.shape[0] != positions.shape[0]:
@@ -1469,6 +1471,8 @@ class DeepseekV2Model(nn.Module):
hidden_states, residual = combined_states.split( hidden_states, residual = combined_states.split(
[self.hidden_size, self.hidden_size], dim=-1 [self.hidden_size, self.hidden_size], dim=-1
) )
# fused_add_rms_norm requires a contiguous residual
residual = residual.contiguous()
if self.end_layer in self.aux_hidden_state_layers: if self.end_layer in self.aux_hidden_state_layers:
aux_hidden_states.append(hidden_states + residual) aux_hidden_states.append(hidden_states + residual)
+3
View File
@@ -761,6 +761,7 @@ class Gemma3ForConditionalGeneration(
max_frames_per_batch: int, max_frames_per_batch: int,
device: torch.device, device: torch.device,
dtype: torch.dtype, dtype: torch.dtype,
path: str = "default",
): ):
from vllm.v1.worker.encoder_cudagraph_defs import ( from vllm.v1.worker.encoder_cudagraph_defs import (
EncoderCudaGraphCaptureInputs, EncoderCudaGraphCaptureInputs,
@@ -792,6 +793,7 @@ class Gemma3ForConditionalGeneration(
mm_kwargs: dict[str, Any], mm_kwargs: dict[str, Any],
max_batch_size: int, max_batch_size: int,
max_frames_per_batch: int, max_frames_per_batch: int,
path: str = "default",
): ):
from vllm.v1.worker.encoder_cudagraph_defs import ( from vllm.v1.worker.encoder_cudagraph_defs import (
EncoderCudaGraphReplayBuffers, EncoderCudaGraphReplayBuffers,
@@ -804,6 +806,7 @@ class Gemma3ForConditionalGeneration(
def encoder_cudagraph_forward( def encoder_cudagraph_forward(
self, self,
values: dict[str, torch.Tensor], values: dict[str, torch.Tensor],
path: str = "default",
) -> torch.Tensor: ) -> torch.Tensor:
pixel_values = values["pixel_values"] pixel_values = values["pixel_values"]
image_features = self.vision_tower(pixel_values) image_features = self.vision_tower(pixel_values)