[CI][ROCm] Reduce V1 attention test runtime (#49916)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
@@ -364,22 +364,6 @@ steps:
|
||||
commands:
|
||||
- pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
|
||||
|
||||
- label: V1 attention (H100-MI250) # TBD
|
||||
timeout_in_minutes: 180
|
||||
mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
|
||||
agent_pool: mi250_1
|
||||
working_dir: "/vllm-workspace/tests"
|
||||
source_file_dependencies:
|
||||
- vllm/config/attention.py
|
||||
- vllm/model_executor/layers/attention
|
||||
- vllm/v1/attention
|
||||
- tests/v1/attention
|
||||
- vllm/_aiter_ops.py
|
||||
- vllm/envs.py
|
||||
- vllm/platforms/rocm.py
|
||||
commands:
|
||||
- pytest -v -s v1/attention
|
||||
|
||||
- label: V1 others (CPU) # TBD
|
||||
timeout_in_minutes: 180
|
||||
mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
|
||||
@@ -2694,11 +2678,12 @@ steps:
|
||||
- export VLLM_WORKER_MULTIPROC_METHOD=spawn
|
||||
- pytest -v -s v1/kv_connector/extract_hidden_states_integration
|
||||
|
||||
- label: V1 attention (H100-MI300) # TBD
|
||||
- label: V1 attention (H100-MI300) %N # TBD
|
||||
timeout_in_minutes: 180
|
||||
mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
|
||||
dind: false
|
||||
agent_pool: mi300_1
|
||||
parallelism: 2
|
||||
optional: true
|
||||
working_dir: "/vllm-workspace/tests"
|
||||
source_file_dependencies:
|
||||
@@ -2710,7 +2695,7 @@ steps:
|
||||
- vllm/envs.py
|
||||
- vllm/platforms/rocm.py
|
||||
commands:
|
||||
- pytest -v -s v1/attention
|
||||
- pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
|
||||
|
||||
- label: V1 Core + KV + Metrics # TBD
|
||||
timeout_in_minutes: 180
|
||||
@@ -3676,11 +3661,12 @@ steps:
|
||||
|
||||
#------------------------------------------------------------ mi355 · v1 -------------------------------------------------------------#
|
||||
|
||||
- label: V1 attention (B200-MI355) # TBD
|
||||
- label: V1 attention (B200-MI355) %N # TBD
|
||||
timeout_in_minutes: 180
|
||||
mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
|
||||
dind: false
|
||||
agent_pool: mi355_1
|
||||
parallelism: 2
|
||||
working_dir: "/vllm-workspace/tests"
|
||||
source_file_dependencies:
|
||||
- vllm/config/attention.py
|
||||
@@ -3691,7 +3677,7 @@ steps:
|
||||
- vllm/envs.py
|
||||
- vllm/platforms/rocm.py
|
||||
commands:
|
||||
- pytest -v -s v1/attention
|
||||
- pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
|
||||
|
||||
- label: V1 Core + KV + Metrics # TBD
|
||||
timeout_in_minutes: 180
|
||||
|
||||
@@ -40,6 +40,10 @@ from vllm.v1.attention.backends.mla.prefill import (
|
||||
MLAPrefillBackendEnum,
|
||||
get_mla_prefill_backend,
|
||||
)
|
||||
from vllm.v1.attention.backends.mla.prefill.base import MLADimensions
|
||||
from vllm.v1.attention.backends.mla.prefill.selector import (
|
||||
MLAPrefillSelectorConfig,
|
||||
)
|
||||
from vllm.v1.attention.backends.registry import AttentionBackendEnum
|
||||
from vllm.v1.attention.ops.flashmla import is_flashmla_dense_supported
|
||||
from vllm.v1.kv_cache_interface import (
|
||||
@@ -145,14 +149,67 @@ def test_mla_post_load_preserves_runtime_weight_addresses(monkeypatch):
|
||||
torch.testing.assert_close(layer.W_UK_T, old_w_uk_t + 100)
|
||||
|
||||
|
||||
# Filtered per-test via validate_configuration (capability/deps/dims).
|
||||
# Validate parameter combinations during collection, before GPU fixtures run.
|
||||
PREFILL_BACKENDS_TO_TEST = [
|
||||
MLAPrefillBackendEnum.ROCM_AITER_FA,
|
||||
MLAPrefillBackendEnum.FLASH_ATTN,
|
||||
MLAPrefillBackendEnum.FLASHINFER,
|
||||
MLAPrefillBackendEnum.TRTLLM_RAGGED,
|
||||
MLAPrefillBackendEnum.TOKENSPEED_MLA,
|
||||
]
|
||||
|
||||
MLA_DIMENSIONS_TO_TEST = [
|
||||
("deepseek", 128, 128),
|
||||
("glm", 192, 256),
|
||||
]
|
||||
|
||||
|
||||
def _prefill_backend_dimension_params():
|
||||
device_capability = current_platform.get_device_capability()
|
||||
params = []
|
||||
for prefill_backend in PREFILL_BACKENDS_TO_TEST:
|
||||
for dimensions_id, qk_nope_head_dim, v_head_dim in MLA_DIMENSIONS_TO_TEST:
|
||||
if device_capability is None:
|
||||
invalid_reasons = ["device capability unavailable"]
|
||||
else:
|
||||
try:
|
||||
invalid_reasons = (
|
||||
prefill_backend.get_class().validate_configuration(
|
||||
device_capability,
|
||||
MLAPrefillSelectorConfig(
|
||||
dtype=torch.bfloat16,
|
||||
mla_dimensions=MLADimensions(
|
||||
qk_nope_head_dim=qk_nope_head_dim,
|
||||
qk_rope_head_dim=64,
|
||||
v_head_dim=v_head_dim,
|
||||
),
|
||||
),
|
||||
)
|
||||
)
|
||||
except ImportError:
|
||||
invalid_reasons = ["ImportError"]
|
||||
|
||||
marks = []
|
||||
if invalid_reasons:
|
||||
marks.append(
|
||||
pytest.mark.skip(
|
||||
reason=(
|
||||
f"Prefill backend {prefill_backend.name} unavailable: "
|
||||
f"{invalid_reasons}"
|
||||
)
|
||||
)
|
||||
)
|
||||
params.append(
|
||||
pytest.param(
|
||||
prefill_backend,
|
||||
qk_nope_head_dim,
|
||||
v_head_dim,
|
||||
id=f"{dimensions_id}-{prefill_backend}",
|
||||
marks=marks,
|
||||
)
|
||||
)
|
||||
return params
|
||||
|
||||
|
||||
SPEC_DECODE_BACKENDS = []
|
||||
for backend in BACKENDS_TO_TEST:
|
||||
@@ -1098,11 +1155,9 @@ def run_attention_backend(
|
||||
@pytest.mark.parametrize("tensor_parallel_size", [1, 4, 8, 16])
|
||||
@pytest.mark.parametrize("kv_cache_dtype", ["auto", "fp8", "fp8_e4m3"])
|
||||
@pytest.mark.parametrize(("q_scale", "k_scale"), [(1.0, 1.0), (2.0, 3.0)])
|
||||
@pytest.mark.parametrize("prefill_backend", PREFILL_BACKENDS_TO_TEST)
|
||||
@pytest.mark.parametrize(
|
||||
("qk_nope_head_dim", "v_head_dim"),
|
||||
[(128, 128), (192, 256)],
|
||||
ids=["deepseek", "glm"],
|
||||
("prefill_backend", "qk_nope_head_dim", "v_head_dim"),
|
||||
_prefill_backend_dimension_params(),
|
||||
)
|
||||
def test_backend_correctness(
|
||||
default_vllm_config,
|
||||
@@ -1153,32 +1208,6 @@ def test_backend_correctness(
|
||||
if not backends_to_test:
|
||||
pytest.skip(f"No backends support kv_cache_dtype={kv_cache_dtype}")
|
||||
|
||||
# Skip prefill backends that can't satisfy capability/deps/dimension constraints.
|
||||
from vllm.v1.attention.backends.mla.prefill.base import MLADimensions
|
||||
from vllm.v1.attention.backends.mla.prefill.selector import (
|
||||
MLAPrefillSelectorConfig,
|
||||
)
|
||||
|
||||
try:
|
||||
prefill_invalid_reasons = prefill_backend.get_class().validate_configuration(
|
||||
current_platform.get_device_capability(),
|
||||
MLAPrefillSelectorConfig(
|
||||
dtype=torch.bfloat16,
|
||||
mla_dimensions=MLADimensions(
|
||||
qk_nope_head_dim=qk_nope_head_dim,
|
||||
qk_rope_head_dim=64,
|
||||
v_head_dim=v_head_dim,
|
||||
),
|
||||
),
|
||||
)
|
||||
except ImportError:
|
||||
prefill_invalid_reasons = ["ImportError"]
|
||||
if prefill_invalid_reasons:
|
||||
pytest.skip(
|
||||
f"Prefill backend {prefill_backend.name} unavailable: "
|
||||
f"{prefill_invalid_reasons}"
|
||||
)
|
||||
|
||||
batch_spec = BATCH_SPECS[batch_spec_name]
|
||||
is_spec_decode_test = batch_spec_name.startswith("spec_decode")
|
||||
unique_block_sizes = sorted(set(BACKEND_BLOCK_SIZES[b] for b in backends_to_test))
|
||||
|
||||
Reference in New Issue
Block a user