[ROCm][Bugfix] Register MRV2 sampler JIT warmups (#58092)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Andreas Karatzas
2026-09-23 03:17:13 -05:00
committed by GitHub
co-authored by OpenAI Codex
parent 26879f3260
commit 96e203b08e
4 changed files with 114 additions and 0 deletions
+24
View File
@@ -5252,3 +5252,27 @@ steps:
- amdgfx950nightly
- amdmi355
optional: true
- label: ":amd: (MI355) MRV2 Sampler JIT Warmup"
timeout_in_minutes: 10
mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
dind: false
agent_pool: mi355_1
num_gpus: 1
optional: true
working_dir: "/vllm-workspace/"
source_file_dependencies:
- tests/jit_monitor/
- tests/models/utils.py
- tests/models/registry.py
- tests/utils.py
- vllm/v1/worker/gpu/
- vllm/v1/worker/gpu_worker.py
- vllm/model_executor/warmup/
- vllm/v1/sample/ops/topk_topp_sampler.py
- vllm/utils/jit_monitor.py
- vllm/config/observability.py
- vllm/platforms/rocm.py
- vllm/envs.py
commands:
- pytest -v -s tests/jit_monitor/test_no_runtime_jit_rocm.py --timeout=300
+21
View File
@@ -20,3 +20,24 @@ steps:
# instead of running until the build timeout.
- export PYTHONFAULTHANDLER=1
- pytest -v -s jit_monitor/test_no_runtime_jit.py --timeout=900 --timeout-method=thread
mirror:
amd:
# Focused MRV2 regression; the broader model/backend matrix is pending.
label: ':amd: (MI355) MRV2 Sampler JIT Warmup'
dind: false
device: mi355_1
num_devices: 1
timeout_in_minutes: 10
working_dir: /vllm-workspace/
depends_on:
- image-build-amd
source_file_dependencies:
- tests/jit_monitor/
- tests/models/utils.py
- tests/utils.py
- vllm/v1/worker/gpu/
- vllm/v1/sample/ops/topk_topp_sampler.py
- vllm/platforms/rocm.py
- vllm/envs.py
commands:
- pytest -v -s tests/jit_monitor/test_no_runtime_jit_rocm.py --timeout=300
@@ -0,0 +1,60 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Keep the MRV2 sampler warmup registry populated on ROCm."""
import pytest
from vllm import LLM
from vllm.platforms import current_platform
from ..models.utils import dummy_hf_overrides
from ..utils import create_new_process_for_each_test
from .test_no_runtime_jit import _run_shape_battery
pytestmark = pytest.mark.skipif(not current_platform.is_rocm(), reason="Requires ROCm")
def _worker_monitor_state(worker):
from vllm.utils import jit_monitor
return (
jit_monitor.is_active(),
worker.observability_config.jit_monitor_mode,
)
@create_new_process_for_each_test("spawn")
def _run_rocm_shape_battery() -> None:
llm = LLM(
"Qwen/Qwen3-0.6B",
max_model_len=2048,
max_num_seqs=8,
gpu_memory_utilization=0.03,
kv_cache_memory_bytes=256 * 1024 * 1024,
load_format="dummy",
hf_overrides=dummy_hf_overrides,
enforce_eager=False,
jit_monitor_mode="error",
)
try:
states = llm.collective_rpc(_worker_monitor_state, timeout=30)
assert states and all(state == (True, "error") for state in states)
_run_shape_battery(llm)
finally:
llm.llm_engine.engine_core.shutdown(timeout=30)
def test_v2_sampler_warmup_rocm(monkeypatch, tmp_path):
"""Warmup must cover the existing prefill, decode, and sampler battery.
A fresh cache and child process prevent previous model executions from
hiding a missing warmup specialization.
"""
monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1")
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "0")
monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
monkeypatch.setenv("TRITON_CACHE_DIR", str(tmp_path / "triton"))
monkeypatch.setenv("VLLM_CACHE_ROOT", str(tmp_path / "vllm"))
monkeypatch.setenv("TORCHINDUCTOR_CACHE_DIR", str(tmp_path / "inductor"))
_run_rocm_shape_battery()
+9
View File
@@ -442,6 +442,15 @@ class GPUModelRunner(LoRAModelRunnerMixin):
self.vllm_config, self.req_states, self.is_pooling_model, config_processors
)
if self.is_last_pp_rank and not self.is_pooling_model:
from vllm.v1.sample.ops.topk_topp_sampler import (
register_top_k_top_p_warmups,
)
# V2 bypasses TopKTopPSampler, which registers native warmups.
# CUDA also needs these for its FlashInfer fallback paths.
with self.jit_warmup_registry.activate():
register_top_k_top_p_warmups()
sampler_kwargs: dict[str, Any] = {
"vllm_config": self.vllm_config,
"max_num_reqs": self.max_num_reqs,