Merge remote-tracking branch 'upstream/main' into patches

This commit is contained in:
2026-09-28 12:06:52 +07:00
170 changed files with 7605 additions and 1142 deletions
@@ -49,6 +49,7 @@ PRUNED_METADATA_FIELDS = {
"prefill_state_indices",
"prefill_has_initial_state",
"spec_sequence_masks",
"spec_sequence_masks_cpu",
"uniform_spec_sequence_length",
"flashinfer_prefill_query_start_loc",
"flashinfer_prefill_seq_order",
@@ -175,6 +175,24 @@ MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
},
marks=[pytest.mark.core_model],
),
"minimax_m3": VitCudagraphTestConfig(
model="MiniMaxAI/MiniMax-M3",
modalities=["image"],
image_prompt=("]~b]user\n]<]image[>[What is in this image?[e~[\n]~b]ai\n"),
compilation_config_overrides={
"encoder_cudagraph_token_budgets": [512],
},
vllm_runner_kwargs={
"trust_remote_code": True,
"load_format": "dummy",
"hf_overrides": partial(
dummy_hf_overrides,
model_arch="MiniMaxM3SparseForConditionalGeneration",
),
"mm_processor_kwargs": {"max_long_side_pixel": 336},
},
marks=[pytest.mark.core_model],
),
"qwen3_vl": VitCudagraphTestConfig(
model="Qwen/Qwen3-VL-2B-Instruct",
image_prompt=qwen_vl_chat_template(
+3 -18
View File
@@ -5,13 +5,7 @@
import pytest
import torch
from tests.utils import ensure_current_vllm_config
from vllm.distributed.parallel_state import (
init_distributed_environment,
initialize_model_parallel,
)
from vllm.platforms import current_platform
from vllm.utils.network_utils import get_open_port
EMBED_DIM = 256
NUM_HEADS = 4
@@ -21,20 +15,11 @@ WINDOW = 8
SEQ_LENS = [5, 37]
@pytest.fixture(scope="module")
def vision_attn_env():
init_distributed_environment(
world_size=1,
rank=0,
local_rank=0,
distributed_init_method=f"tcp://127.0.0.1:{get_open_port()}",
backend="nccl",
)
@pytest.fixture
def vision_attn_env(dist_init):
default_dtype = torch.get_default_dtype()
torch.set_default_dtype(torch.bfloat16)
with ensure_current_vllm_config():
initialize_model_parallel(tensor_model_parallel_size=1)
yield
yield
torch.set_default_dtype(default_dtype)