[ROCm][CI] Add MI355 TP2 AR-RMS and B200 fusion mirrors (#58284)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Andreas Karatzas
2026-09-29 07:10:26 +00:00
committed by GitHub
co-authored by OpenAI Codex
parent 69d5c81fcd
commit 78fc1e6274
4 changed files with 165 additions and 1 deletions
+50
View File
@@ -3228,6 +3228,56 @@ steps:
commands:
- pytest -v -s compile/h100/test_startup.py
- label: ":amd: (MI355) Fusion E2E TP2 AR-RMS FP8 Config Sweep"
timeout_in_minutes: 45
mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
dind: false
agent_pool: mi355_2
num_gpus: 2
optional: true
working_dir: "/vllm-workspace/"
source_file_dependencies:
- vllm/compilation/
- vllm/model_executor/
- vllm/distributed/
- vllm/v1/attention/
- vllm/_aiter_ops.py
- vllm/platforms/rocm.py
- tests/compile/fusions_e2e/
- tests/utils.py
- tests/conftest.py
- csrc/quantization/
commands:
# Existing MI300 jobs own BF16; TP2 Quick owns the excluded Llama cell.
- amd-smi
- VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_CUSTOM_AR=1 pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k 'test_tp2_ar_rms_fp8_fusions and Llama-3 and ROCM_AITER_UNIFIED_ATTN and (dynamo_partition or +quant_fp8 or +rms_norm)'
- label: ":amd: (MI355) Fusion E2E TP2 FP8 BF16 rewrites"
timeout_in_minutes: 45
mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
dind: false
agent_pool: mi355_2
num_gpus: 2
optional: true
working_dir: "/vllm-workspace/"
source_file_dependencies:
- vllm/compilation/
- vllm/model_executor/
- vllm/distributed/
- vllm/v1/attention/
- vllm/_aiter_ops.py
- vllm/platforms/rocm.py
- tests/compile/fusions_e2e/
- tests/utils.py
- tests/conftest.py
- csrc/quantization/
commands:
# Dummy-weight rewrite coverage; trained FP8 AR/RMS equivalence is unresolved.
# TODO: FP4 output-quant and trained MXFP4 AR/RMS fusion remain unsupported.
- amd-smi
- VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_CUSTOM_AR=1 pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k 'test_tp2_ar_rms_fp8_fusions and qwen3 and ROCM_AITER_UNIFIED_ATTN and +quant_fp8 and (dynamo_partition or not +rms_norm)'
- pytest -v -s tests/compile/fusions_e2e/test_tp2_rocm_moe.py
- label: ":amd: (MI355) Fusion and Compile FP8"
timeout_in_minutes: 45
mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+52
View File
@@ -267,6 +267,31 @@ steps:
- nvidia-smi
# Run just llama3 (fp8 & bf16) for all config combinations
- pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "llama-3"
mirror:
amd:
label: ":amd: (MI355) Fusion E2E TP2 AR-RMS FP8 Config Sweep"
dind: false
device: mi355_2
num_devices: 2
timeout_in_minutes: 45
working_dir: "/vllm-workspace/"
depends_on:
- image-build-amd
source_file_dependencies:
- vllm/compilation/
- vllm/model_executor/
- vllm/distributed/
- vllm/v1/attention/
- vllm/_aiter_ops.py
- vllm/platforms/rocm.py
- tests/compile/fusions_e2e/
- tests/utils.py
- tests/conftest.py
- csrc/quantization/
commands:
# Existing MI300 jobs own BF16; TP2 Quick owns the excluded Llama cell.
- amd-smi
- VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_CUSTOM_AR=1 pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k 'test_tp2_ar_rms_fp8_fusions and Llama-3 and ROCM_AITER_UNIFIED_ATTN and (dynamo_partition or +quant_fp8 or +rms_norm)'
- label: ":nvidia: (H100) Fusion E2E TP2 AsyncTP Config Sweep"
key: fusion-e2e-tp2-asynctp-config-sweep-h100
@@ -334,3 +359,30 @@ steps:
# for ar-rms-quant-fp4, also sweep llama3
- pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "(FLASHINFER and inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek))) or Llama-3.1-8B-Instruct-FP4"
- pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "FLASHINFER and inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek))"
mirror:
amd:
label: ":amd: (MI355) Fusion E2E TP2 FP8 BF16 rewrites"
dind: false
device: mi355_2
num_devices: 2
timeout_in_minutes: 45
working_dir: "/vllm-workspace/"
depends_on:
- image-build-amd
source_file_dependencies:
- vllm/compilation/
- vllm/model_executor/
- vllm/distributed/
- vllm/v1/attention/
- vllm/_aiter_ops.py
- vllm/platforms/rocm.py
- tests/compile/fusions_e2e/
- tests/utils.py
- tests/conftest.py
- csrc/quantization/
commands:
# Dummy-weight rewrite coverage; trained FP8 AR/RMS equivalence is unresolved.
# TODO: FP4 output-quant and trained MXFP4 AR/RMS fusion remain unsupported.
- amd-smi
- VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_CUSTOM_AR=1 pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k 'test_tp2_ar_rms_fp8_fusions and qwen3 and ROCM_AITER_UNIFIED_ATTN and +quant_fp8 and (dynamo_partition or not +rms_norm)'
- pytest -v -s tests/compile/fusions_e2e/test_tp2_rocm_moe.py
+3 -1
View File
@@ -111,7 +111,9 @@ def test_tp2_ar_rms_fp8_fusions(
model_kwargs["max_model_len"] = 1024
model_kwargs["kernel_config"] = {"enable_flashinfer_autotune": False}
model_kwargs["disable_custom_all_reduce"] = False
if current_platform.is_rocm() and "llama-3" in model_name_lower:
if current_platform.is_rocm() and (
"llama-3" in model_name_lower or "qwen" in model_name_lower
):
model_kwargs.update(
gpu_memory_utilization=0.01,
kv_cache_memory_bytes=64 * 1024**2,
@@ -0,0 +1,60 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""ROCm MoE coverage for the TP2 GEMM/collective fusion contract."""
import pytest
from tests.utils import multi_gpu_test
from vllm.config import PassConfig
from vllm.platforms import current_platform
from .common import INDUCTOR_GRAPH_PARTITION
from .models import TRITON_ATTN, qwen3_a3b
@multi_gpu_test(num_gpus=2)
@pytest.mark.skipif(not current_platform.is_rocm(), reason="ROCm AsyncTP contract")
@pytest.mark.parametrize("inductor_graph_partition", INDUCTOR_GRAPH_PARTITION[:1])
@pytest.mark.parametrize("attn_backend", [TRITON_ATTN])
def test_tp2_rocm_moe_async_tp_bf16_fusions(
inductor_graph_partition, attn_backend, run_e2e_fusion_test
):
"""MoE dense projections retain the Qwen SP/AsyncTP rewrite counts."""
n_layers = 4
run_e2e_fusion_test(
qwen3_a3b.model_name,
qwen3_a3b.matches(n_layers),
dict(
hf_overrides=dict(
num_hidden_layers=n_layers,
hidden_size=512,
intermediate_size=1024,
moe_intermediate_size=128,
num_attention_heads=8,
num_key_value_heads=2,
head_dim=64,
num_experts=8,
num_experts_per_tok=2,
),
load_format="dummy",
max_model_len=1024,
kv_cache_memory_bytes=64 * 1024**2,
gpu_memory_utilization=0.01,
max_num_batched_tokens=1024,
max_num_seqs=32,
),
attn_backend,
dict(
use_inductor_graph_partition=inductor_graph_partition,
custom_ops=["-rms_norm"],
pass_config=PassConfig(
enable_qk_norm_rope_fusion=True,
enable_sp=True,
fuse_gemm_comms=True,
fuse_allreduce_rms=False,
sp_min_token_num=512,
),
),
["norm_rope_fusion", "sequence_parallel", "async_tp"],
tp_size=2,
)