[MyPy][3/N] Fix MyPy errors in test groups (part 3) (#55939)
Signed-off-by: Taimys <ffy2281693949@gmail.com> Signed-off-by: Luyi Xiao <81631911+luyixiao95@users.noreply.github.com> Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Codex <noreply@openai.com> Co-authored-by: Luyi Xiao <81631911+luyixiao95@users.noreply.github.com> Co-authored-by: Taimys <ffy2281693949@gmail.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Codex
Luyi Xiao
Taimys
Harry Mellor
Claude Opus 5.5
parent
c4fa6f36a7
commit
5225f1bacb
@@ -2,6 +2,7 @@
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from contextlib import contextmanager
|
||||
from typing import Literal
|
||||
from unittest.mock import MagicMock, patch
|
||||
|
||||
import pytest
|
||||
@@ -14,6 +15,7 @@ from vllm.config import (
|
||||
VllmConfig,
|
||||
set_current_vllm_config,
|
||||
)
|
||||
from vllm.config.cache import CacheDType
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.platforms.cpu import CpuPlatform
|
||||
from vllm.platforms.interface import DeviceCapability
|
||||
@@ -21,12 +23,12 @@ from vllm.platforms.interface import DeviceCapability
|
||||
if current_platform.is_cuda():
|
||||
from vllm.platforms.cuda import CudaPlatform
|
||||
else:
|
||||
CudaPlatform = None
|
||||
CudaPlatform = None # type: ignore[assignment] # Unavailable platform import.
|
||||
|
||||
if current_platform.is_rocm():
|
||||
from vllm.platforms.rocm import RocmPlatform
|
||||
else:
|
||||
RocmPlatform = None
|
||||
RocmPlatform = None # type: ignore[misc, assignment] # Unavailable platform import.
|
||||
|
||||
from vllm.v1.attention.backend import AttentionType
|
||||
from vllm.v1.attention.backends.registry import AttentionBackendEnum
|
||||
@@ -354,14 +356,14 @@ def test_invalid_backend():
|
||||
def test_auto_backend_string(auto_value: str):
|
||||
"""Test that 'auto' string value triggers automatic backend selection."""
|
||||
# Using "auto" should result in backend=None (automatic selection)
|
||||
attention_config = AttentionConfig(backend=auto_value)
|
||||
attention_config = AttentionConfig(backend=auto_value) # type: ignore[arg-type]
|
||||
assert attention_config.backend is None
|
||||
|
||||
|
||||
def test_auto_backend_selection_behavior():
|
||||
"""Test that 'auto' backend behaves same as None (automatic selection)."""
|
||||
# Create config with explicit "auto"
|
||||
auto_config = AttentionConfig(backend="auto")
|
||||
auto_config = AttentionConfig(backend="auto") # type: ignore[arg-type]
|
||||
|
||||
# Create config with None (default)
|
||||
none_config = AttentionConfig(backend=None)
|
||||
@@ -406,7 +408,7 @@ def test_auto_backend_selection_behavior():
|
||||
reason="Attention backend FA3 is not supported on ROCm. This test can't succeed.",
|
||||
)
|
||||
def test_per_head_quant_scales_backend_selection(
|
||||
backend_name: str, flash_attn_version: int | None, should_succeed: bool
|
||||
backend_name: str, flash_attn_version: Literal[2, 3, 4] | None, should_succeed: bool
|
||||
):
|
||||
"""Test backend selection when use_per_head_quant_scales=True."""
|
||||
# Clear cache to ensure fresh backend selection
|
||||
@@ -559,7 +561,7 @@ def test_non_causal_autoselect_backend():
|
||||
"int8_per_token_head",
|
||||
],
|
||||
)
|
||||
def test_flash_attn_rejects_unhandled_kv_cache_dtypes(kv_cache_dtype: str):
|
||||
def test_flash_attn_rejects_unhandled_kv_cache_dtypes(kv_cache_dtype: CacheDType):
|
||||
"""FlashAttentionBackend must not claim support for kv_cache dtypes
|
||||
that it cannot handle."""
|
||||
from vllm.v1.attention.backends.flash_attn import FlashAttentionBackend
|
||||
@@ -569,7 +571,7 @@ def test_flash_attn_rejects_unhandled_kv_cache_dtypes(kv_cache_dtype: str):
|
||||
|
||||
@pytest.mark.parametrize("kv_cache_dtype", ["fp8", "fp8_e4m3"])
|
||||
def test_flash_attn_accepts_handled_fp8_variants(
|
||||
kv_cache_dtype: str, monkeypatch: pytest.MonkeyPatch
|
||||
kv_cache_dtype: CacheDType, monkeypatch: pytest.MonkeyPatch
|
||||
):
|
||||
"""FlashAttentionBackend must accept the two fp8 dtypes it can actually
|
||||
handle: 'fp8' (alias for fp8_e4m3fn) and 'fp8_e4m3'."""
|
||||
@@ -876,7 +878,9 @@ def test_rswa_selection_does_not_reuse_causal_result(blackwell_selection):
|
||||
config = EngineArgs(
|
||||
model="google/gemma-4-31B-it",
|
||||
dtype="bfloat16",
|
||||
attention_config={"backend": "TRITON_FLASHINFER"},
|
||||
attention_config=AttentionConfig(
|
||||
backend=AttentionBackendEnum.TRITON_FLASHINFER
|
||||
),
|
||||
).create_engine_config()
|
||||
with set_current_vllm_config(config):
|
||||
assert (
|
||||
|
||||
@@ -151,7 +151,7 @@ def test_cascade(
|
||||
max_seqlen_k=max_kv_len,
|
||||
softmax_scale=scale,
|
||||
causal=True,
|
||||
window_size=window_size,
|
||||
window_size=list(window_size),
|
||||
block_table=block_tables,
|
||||
softcap=soft_cap if soft_cap is not None else 0,
|
||||
)
|
||||
|
||||
@@ -836,7 +836,7 @@ def varlen_encoder_zentorch_sdpa(
|
||||
key,
|
||||
value,
|
||||
output,
|
||||
SimpleNamespace(query_start_loc=query_start_loc, causal=False),
|
||||
SimpleNamespace(query_start_loc=query_start_loc, causal=False), # type: ignore[arg-type]
|
||||
scale,
|
||||
sliding_window if sliding_window is not None else -1,
|
||||
alibi_slopes,
|
||||
|
||||
@@ -187,7 +187,7 @@ def test_varlen_with_paged_kv(
|
||||
max_seqlen_k=max_kv_len,
|
||||
softmax_scale=scale,
|
||||
causal=True,
|
||||
window_size=window_size,
|
||||
window_size=list(window_size),
|
||||
block_table=block_tables,
|
||||
softcap=soft_cap if soft_cap is not None else 0,
|
||||
fa_version=fa_version,
|
||||
@@ -274,7 +274,7 @@ def test_fa4_hd256_paged_call_shape(
|
||||
max_seqlen_k=num_pages * block_size,
|
||||
softmax_scale=scale,
|
||||
causal=True,
|
||||
window_size=window_size,
|
||||
window_size=list(window_size),
|
||||
block_table=block_tables[:, :num_pages],
|
||||
num_splits=1,
|
||||
fa_version=4,
|
||||
|
||||
@@ -197,11 +197,6 @@ def test_decode_concat_ignores_negative_slots_for_cache(cache_format: str) -> No
|
||||
mixed_cache = initial_cache.clone()
|
||||
reference_cache = _reference_cache(cache_format, inputs, initial_cache, slots)
|
||||
scale_inv = torch.ones(1, device="cuda", dtype=torch.float32)
|
||||
kwargs = {
|
||||
"ds_mla": cache_format == "fp8_ds_mla",
|
||||
"q_scale_inv": scale_inv if cache_format == "fp8" else None,
|
||||
"cache_scale_inv": scale_inv if cache_format == "fp8" else None,
|
||||
}
|
||||
|
||||
output = fused_mla_decode_q_concat_kv_cache_insert(
|
||||
inputs["ql_nope"],
|
||||
@@ -210,7 +205,9 @@ def test_decode_concat_ignores_negative_slots_for_cache(cache_format: str) -> No
|
||||
inputs["k_pe"],
|
||||
mixed_cache,
|
||||
slots,
|
||||
**kwargs,
|
||||
ds_mla=cache_format == "fp8_ds_mla",
|
||||
q_scale_inv=scale_inv if cache_format == "fp8" else None,
|
||||
cache_scale_inv=scale_inv if cache_format == "fp8" else None,
|
||||
)
|
||||
|
||||
expected = _latent_query(inputs["ql_nope"], inputs["q_pe"])
|
||||
|
||||
@@ -277,7 +277,7 @@ def test_mha_attn_varlen_forward_flashinfer(
|
||||
# Override vllm config so get_vit_attn_backend returns FLASHINFER (simulates
|
||||
# --mm-encoder-attn-backend=FLASHINFER).
|
||||
vllm_config = get_current_vllm_config()
|
||||
old_model_config = getattr(vllm_config, "model_config", None)
|
||||
old_model_config = vllm_config.model_config
|
||||
minimal_model_config = type(
|
||||
"MinimalModelConfig",
|
||||
(),
|
||||
@@ -376,7 +376,7 @@ def test_mha_attn_varlen_forward_aiter_fp8(
|
||||
torch.set_default_dtype(dtype)
|
||||
|
||||
vllm_config = get_current_vllm_config()
|
||||
old_model_config = getattr(vllm_config, "model_config", None)
|
||||
old_model_config = vllm_config.model_config
|
||||
minimal_model_config = type(
|
||||
"MinimalModelConfig",
|
||||
(),
|
||||
@@ -442,7 +442,7 @@ def test_mha_attn_varlen_forward_aiter_fp8(
|
||||
)
|
||||
def test_mha_attn_aiter_fp8_rejects_unsupported_arch(default_vllm_config):
|
||||
vllm_config = get_current_vllm_config()
|
||||
old_model_config = getattr(vllm_config, "model_config", None)
|
||||
old_model_config = vllm_config.model_config
|
||||
vllm_config.model_config = type(
|
||||
"MinimalModelConfig",
|
||||
(),
|
||||
@@ -468,7 +468,7 @@ def test_mha_attn_aiter_fp8_rejects_unsupported_arch(default_vllm_config):
|
||||
)
|
||||
def test_mha_attn_fp8_rejects_wrong_backend(default_vllm_config):
|
||||
vllm_config = get_current_vllm_config()
|
||||
old_model_config = getattr(vllm_config, "model_config", None)
|
||||
old_model_config = vllm_config.model_config
|
||||
vllm_config.model_config = type(
|
||||
"MinimalModelConfig",
|
||||
(),
|
||||
|
||||
@@ -1689,8 +1689,10 @@ def test_aiter_sparse_pa_layout_contract(monkeypatch, layout):
|
||||
)
|
||||
|
||||
backend = sparse_attn_mod.MiniMaxM3SparseBackend
|
||||
assert KVCacheLayout.LHBNC in backend.supported_kv_cache_layouts()
|
||||
assert KVCacheLayout.LBHNC in backend.supported_kv_cache_layouts()
|
||||
layouts = backend.supported_kv_cache_layouts()
|
||||
assert layouts is not None
|
||||
assert KVCacheLayout.LHBNC in layouts
|
||||
assert KVCacheLayout.LBHNC in layouts
|
||||
|
||||
nb, h = 3, 1
|
||||
cache_dtype = current_platform.fp8_dtype()
|
||||
@@ -1772,19 +1774,22 @@ def test_aiter_indexer_requires_the_aiter_attend(monkeypatch):
|
||||
import vllm.models.minimax_m3.amd.indexer_aiter as indexer_aiter_mod
|
||||
|
||||
monkeypatch.setattr(indexer_aiter_mod.current_platform, "is_rocm", lambda: True)
|
||||
kwargs = dict(
|
||||
topk_blocks=TOPK,
|
||||
sparse_block_size=BLOCK_SIZE,
|
||||
num_index_heads=1,
|
||||
index_head_dim=HEAD_DIM,
|
||||
indexer_kv_dtype="fp8_e4m3",
|
||||
max_model_len=8192,
|
||||
)
|
||||
|
||||
def unsupported_reason() -> str | None:
|
||||
return indexer_aiter_mod.aiter_indexer_unsupported_reason(
|
||||
topk_blocks=TOPK,
|
||||
sparse_block_size=BLOCK_SIZE,
|
||||
num_index_heads=1,
|
||||
index_head_dim=HEAD_DIM,
|
||||
# The implementation also accepts this legacy fp8 alias.
|
||||
indexer_kv_dtype="fp8_e4m3", # type: ignore[arg-type]
|
||||
max_model_len=8192,
|
||||
)
|
||||
|
||||
monkeypatch.setattr(
|
||||
indexer_aiter_mod, "_minimax_m3_aiter_sparse_pa_requested", lambda: False
|
||||
)
|
||||
reason = indexer_aiter_mod.aiter_indexer_unsupported_reason(**kwargs)
|
||||
reason = unsupported_reason()
|
||||
assert reason is not None and "AITER sparse PA attend" in reason
|
||||
|
||||
# With the attend asked for, the gate moves on to the kernel-contract
|
||||
@@ -1792,7 +1797,7 @@ def test_aiter_indexer_requires_the_aiter_attend(monkeypatch):
|
||||
monkeypatch.setattr(
|
||||
indexer_aiter_mod, "_minimax_m3_aiter_sparse_pa_requested", lambda: True
|
||||
)
|
||||
reason = indexer_aiter_mod.aiter_indexer_unsupported_reason(**kwargs)
|
||||
reason = unsupported_reason()
|
||||
assert reason is None or "AITER sparse PA attend" not in reason
|
||||
|
||||
|
||||
|
||||
@@ -235,14 +235,14 @@ def test_msa_metadata_builder_prepares_cutlass_for_regular_decode(
|
||||
builder.num_q_heads = 64
|
||||
builder.num_kv_heads = 4
|
||||
builder.topk_blocks = TOPK
|
||||
builder.kv_cache_spec = SimpleNamespace(num_kv_heads=4)
|
||||
builder.kv_cache_spec = SimpleNamespace(num_kv_heads=4) # type: ignore[assignment]
|
||||
builder.kv_cache_dtype = "fp8_e4m3"
|
||||
builder.decode_backend = "cutlass"
|
||||
builder.msa_cutlass_plan_cache = object()
|
||||
builder.msa_cutlass_plan_cache = object() # type: ignore[assignment] # Mocked plan-cache sentinel.
|
||||
|
||||
metadata = builder.build(
|
||||
0,
|
||||
SimpleNamespace(
|
||||
SimpleNamespace( # type: ignore[arg-type]
|
||||
seq_lens_cpu_upper_bound=torch.full((batch,), 257, dtype=torch.int32)
|
||||
),
|
||||
)
|
||||
@@ -377,7 +377,7 @@ def test_query_fp8_stays_valid_when_cutlass_plan_appears_on_replay(
|
||||
with torch.cuda.stream(stream), override_forward_context(forward_context):
|
||||
capture = BreakableCUDAGraphCapture()
|
||||
with capture:
|
||||
query_fp8 = MiniMaxM3SparseAttention._allocate_query_fp8(attention, qkv)
|
||||
query_fp8 = MiniMaxM3SparseAttention._allocate_query_fp8(attention, qkv) # type: ignore[arg-type]
|
||||
if query_fp8 is not None:
|
||||
query_fp8.zero_()
|
||||
run_attention(query_fp8)
|
||||
|
||||
@@ -783,6 +783,9 @@ def test_aiter_fa_shared_kv_matches_reference(
|
||||
32 * BLOCK_SIZE, num_kv_heads, head_size, dtype=torch.bfloat16
|
||||
)
|
||||
value = torch.randn_like(key)
|
||||
from vllm.v1.attention.backends.rocm_aiter_fa import AiterFlashAttentionImpl
|
||||
|
||||
assert isinstance(target.impl, AiterFlashAttentionImpl)
|
||||
target.impl.do_kv_cache_update(
|
||||
target,
|
||||
key,
|
||||
|
||||
@@ -82,6 +82,7 @@ def _run_verify_block():
|
||||
)
|
||||
from vllm.config import SpeculativeConfig
|
||||
from vllm.config.vllm import set_current_vllm_config
|
||||
from vllm.v1.attention.backends.mla.rocm_aiter_mla import AiterMLAImpl
|
||||
from vllm.v1.attention.backends.registry import AttentionBackendEnum
|
||||
from vllm.v1.kv_cache_interface import MLAAttentionSpec
|
||||
from vllm.v1.worker.workspace import init_workspace_manager
|
||||
@@ -106,7 +107,7 @@ def _run_verify_block():
|
||||
vllm_config.speculative_config = SpeculativeConfig(
|
||||
method="ngram", num_speculative_tokens=QLEN - 1
|
||||
)
|
||||
vllm_config.model_config.get_num_attention_heads = types.MethodType(
|
||||
vllm_config.model_config.get_num_attention_heads = types.MethodType( # type: ignore[method-assign]
|
||||
lambda self, parallel_config, arch_config=None: NUM_QUERY_HEADS,
|
||||
vllm_config.model_config,
|
||||
)
|
||||
@@ -202,7 +203,8 @@ def _run_verify_block():
|
||||
lambda: spy,
|
||||
),
|
||||
):
|
||||
impl.forward_mqa((q_nope, q_pe), kv_cache, metadata, layer=None)
|
||||
assert isinstance(impl, AiterMLAImpl)
|
||||
impl.forward_mqa((q_nope, q_pe), kv_cache, metadata, layer=None) # type: ignore[arg-type] # Spy decode path does not read layer.
|
||||
|
||||
return metadata, captured
|
||||
|
||||
|
||||
@@ -121,7 +121,7 @@ def _build_prefill_metadata(
|
||||
|
||||
builder = object.__new__(AiterMLAMetadataBuilder)
|
||||
builder.num_heads = num_heads
|
||||
builder.mla_dims = SimpleNamespace(v_head_dim=V_HEAD_DIM)
|
||||
builder.mla_dims = SimpleNamespace(v_head_dim=V_HEAD_DIM) # type: ignore[assignment]
|
||||
builder._init_fp8_prefill_ps_buffers(
|
||||
max_num_reqs=len(seq_lens),
|
||||
max_prefill_qlen=max_q,
|
||||
@@ -136,7 +136,7 @@ def _build_prefill_metadata(
|
||||
prefill = SimpleNamespace(query_start_loc=qo_indptr, max_query_len=max_q)
|
||||
metadata = SimpleNamespace(prefill=prefill, num_decodes=0)
|
||||
common = SimpleNamespace(query_start_loc_cpu=qo_indptr_cpu)
|
||||
builder._build_fp8_prefill_ps_metadata(metadata, common)
|
||||
builder._build_fp8_prefill_ps_metadata(metadata, common) # type: ignore[arg-type]
|
||||
return metadata, total_q, builder
|
||||
|
||||
|
||||
|
||||
@@ -167,7 +167,10 @@ def test_sparse_mla_sink_matches_ragged_reference(
|
||||
impl.sinks = sinks
|
||||
|
||||
output, lse = impl._forward_mla(
|
||||
SimpleNamespace(_q_scale=q_scale, _k_scale=kv_scale), q, kv, metadata
|
||||
SimpleNamespace(_q_scale=q_scale, _k_scale=kv_scale),
|
||||
q,
|
||||
kv,
|
||||
metadata, # type: ignore[arg-type]
|
||||
)
|
||||
kv_flat = kv_ref[:, 0]
|
||||
references = []
|
||||
@@ -241,7 +244,7 @@ def test_sparse_mla_sink_rejects_unsupported_aiter_dtypes(
|
||||
SimpleNamespace(_q_scale=None, _k_scale=None),
|
||||
torch.empty(1, 16, Q_HEAD_DIM, dtype=q_dtype),
|
||||
torch.empty(1, 1, Q_HEAD_DIM, dtype=kv_dtype),
|
||||
metadata,
|
||||
metadata, # type: ignore[arg-type]
|
||||
)
|
||||
|
||||
|
||||
@@ -381,7 +384,7 @@ def test_sparse_mla_sink_matches_dense_attention_with_empty_rows_and_paged_cache
|
||||
paged_kv_indices=torch.tensor(indices, dtype=torch.int32, device="cuda"),
|
||||
paged_kv_indptr=lengths.cumsum(0).to(device="cuda", dtype=torch.int32),
|
||||
)
|
||||
impl = ROCMAiterMLASparseImpl.__new__(ROCMAiterMLASparseImpl)
|
||||
impl = ROCMAiterMLASparseImpl.__new__(ROCMAiterMLASparseImpl) # type: ignore[arg-type]
|
||||
impl.num_heads = num_heads
|
||||
impl.head_size = head_dim
|
||||
impl.kv_lora_rank = value_dim
|
||||
@@ -438,9 +441,9 @@ def test_sparse_mla_backend_resolves_only_contiguous_layer_layouts(monkeypatch,
|
||||
]
|
||||
if layout == "BLNHC":
|
||||
with pytest.raises(ValueError, match="does not satisfy"):
|
||||
resolve_kv_cache_layout(config, supported)
|
||||
resolve_kv_cache_layout(config, supported) # type: ignore[arg-type]
|
||||
else:
|
||||
assert resolve_kv_cache_layout(config, supported).name == layout
|
||||
assert resolve_kv_cache_layout(config, supported).name == layout # type: ignore[arg-type]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16])
|
||||
@@ -491,7 +494,7 @@ def test_sparse_mla_sink_forward_mqa_preserves_split_query(dtype):
|
||||
actual, _ = impl.forward_mqa(
|
||||
(q[..., :V_HEAD_DIM], q[..., V_HEAD_DIM:]),
|
||||
kv,
|
||||
metadata,
|
||||
metadata, # type: ignore[arg-type]
|
||||
SimpleNamespace(_q_scale=None, _k_scale=None),
|
||||
)
|
||||
references = [
|
||||
|
||||
@@ -39,10 +39,10 @@ def _make_builder():
|
||||
topk_tokens = 4
|
||||
|
||||
builder.device = torch.device("cpu")
|
||||
builder.kv_cache_spec = SimpleNamespace(block_size=1)
|
||||
builder.kv_cache_spec = SimpleNamespace(block_size=1) # type: ignore[assignment]
|
||||
builder.model_dtype = torch.bfloat16
|
||||
builder.kv_cache_dtype = "fp8"
|
||||
builder.mla_dims = SimpleNamespace(kv_lora_rank=512, qk_rope_head_dim=64)
|
||||
builder.mla_dims = SimpleNamespace(kv_lora_rank=512, qk_rope_head_dim=64) # type: ignore[assignment]
|
||||
builder.topk_tokens = topk_tokens
|
||||
builder.req_id_per_token_buffer = torch.zeros(
|
||||
max_num_batched_tokens, dtype=torch.int32, device="cpu"
|
||||
|
||||
@@ -227,7 +227,9 @@ def _decode_metadata(case, rows, ratio, query_lens):
|
||||
num_decode_tokens=len(rows),
|
||||
num_prefills=0,
|
||||
num_prefill_tokens=0,
|
||||
decode=types.SimpleNamespace(block_table=block_table, seq_lens=lens[:, None]),
|
||||
decode=types.SimpleNamespace( # type: ignore[arg-type]
|
||||
block_table=block_table, seq_lens=lens[:, None]
|
||||
),
|
||||
decode_row_lens=lens,
|
||||
decode_block_ends=(lens + CAND_BLOCK - 1) // CAND_BLOCK,
|
||||
decode_native=native,
|
||||
@@ -278,7 +280,7 @@ def _prefill_metadata(
|
||||
num_decode_tokens=0,
|
||||
num_prefills=len(case.seq_lens),
|
||||
num_prefill_tokens=len(rows),
|
||||
prefill=types.SimpleNamespace(chunks=chunks),
|
||||
prefill=types.SimpleNamespace(chunks=chunks), # type: ignore[arg-type]
|
||||
prefill_plans=plans,
|
||||
gather_launches=plan_gather_launches(chunks, plans, gather_rows),
|
||||
)
|
||||
|
||||
@@ -955,7 +955,7 @@ def test_rocm_capture_metadata_sets_adaptive_marker(monkeypatch) -> None:
|
||||
)
|
||||
builder = object.__new__(rocm_mod.DeepseekV4ROCMAiterMLASparseMetadataBuilder)
|
||||
|
||||
actual = builder.build_for_cudagraph_capture(SimpleNamespace())
|
||||
actual = builder.build_for_cudagraph_capture(SimpleNamespace()) # type: ignore[arg-type] # Patched base method ignores metadata.
|
||||
|
||||
assert actual is metadata
|
||||
assert actual.for_cudagraph_capture is _on_gfx950()
|
||||
@@ -1563,7 +1563,7 @@ def test_dsv4_adaptive_mla_swa_metadata_graph_replay(monkeypatch) -> None:
|
||||
"sliding_window": window_size,
|
||||
},
|
||||
)
|
||||
vllm_config.speculative_config = SimpleNamespace(
|
||||
vllm_config.speculative_config = SimpleNamespace( # type: ignore[assignment]
|
||||
num_speculative_tokens=upper_query_len - 1,
|
||||
parallel_drafting=False,
|
||||
enable_adaptive_verification=True,
|
||||
|
||||
@@ -153,6 +153,7 @@ def test_triton_unified_attn_diffkv_vs_reference(
|
||||
else:
|
||||
from vllm.v1.attention.backends.fa_utils import flash_attn_varlen_func
|
||||
|
||||
assert fa_version is not None
|
||||
# FA's TMA path needs aligned singleton strides (num_kv_heads == 1).
|
||||
fa_k = canonicalize_singleton_dim_strides(key_cache)
|
||||
fa_v = canonicalize_singleton_dim_strides(value_cache)
|
||||
|
||||
@@ -79,7 +79,7 @@ class TestConfigSet:
|
||||
config_set = ConfigSet("test_kernel")
|
||||
|
||||
with pytest.raises(KeyError, match="platform 'h100' not found"):
|
||||
config_set.get_config("h100", "nonexistent")
|
||||
config_set.get_config("h100", CaseKey({"batch": 32}))
|
||||
|
||||
config_data = {"num_warps": 8, "num_stages": 4}
|
||||
data = {
|
||||
|
||||
@@ -81,7 +81,7 @@ class TestDynamicPerTokenScaledFp8QuantConfigPicker:
|
||||
assert selected_key == CaseKey({"hidden_size": 2048, "num_tokens": 32})
|
||||
|
||||
def test_config_picker_no_configs(self):
|
||||
config_keys: list[dict] = []
|
||||
config_keys: list[CaseKey] = []
|
||||
|
||||
args = _generate_fake_input(16, 4096)
|
||||
selected_key = pick_config(args, config_keys)
|
||||
|
||||
@@ -127,7 +127,7 @@ class TestFusedQkNormRopeConfigPicker:
|
||||
)
|
||||
|
||||
def test_config_picker_no_configs(self):
|
||||
config_keys: list[dict] = []
|
||||
config_keys: list[CaseKey] = []
|
||||
|
||||
args = _generate_fake_input(16, 4096, 128)
|
||||
selected_key = pick_config(args, config_keys)
|
||||
|
||||
@@ -38,6 +38,7 @@ from torch._inductor.pattern_matcher import (
|
||||
)
|
||||
from torch.fx.experimental.proxy_tensor import make_fx
|
||||
|
||||
from vllm.kernels.helion.case_key import CaseKey
|
||||
from vllm.kernels.helion.config_manager import ConfigManager
|
||||
from vllm.kernels.helion.register import HelionKernelWrapper
|
||||
|
||||
@@ -45,7 +46,7 @@ from vllm.kernels.helion.register import HelionKernelWrapper
|
||||
@contextlib.contextmanager
|
||||
def _helion_mock_context():
|
||||
configs = {
|
||||
"default": helion.Config(block_sizes=[64], num_warps=2, num_stages=2),
|
||||
CaseKey.default(): helion.Config(block_sizes=[64], num_warps=2, num_stages=2),
|
||||
}
|
||||
mock_config_manager = Mock(spec=ConfigManager)
|
||||
mock_config_manager.get_platform_configs = Mock(return_value=configs)
|
||||
@@ -88,7 +89,7 @@ class TestMakeFxHop:
|
||||
raw_kernel_func=raw_add_scale,
|
||||
op_name="test_make_fx",
|
||||
fake_impl=lambda *a, **kw: None,
|
||||
config_picker=lambda args, keys: "default",
|
||||
config_picker=lambda args, keys: CaseKey.default(),
|
||||
)
|
||||
|
||||
def fn(x, y):
|
||||
@@ -145,7 +146,7 @@ class TestMakeFxHop:
|
||||
raw_kernel_func=raw_silu_mul,
|
||||
op_name="test_pm_silu_mul",
|
||||
fake_impl=lambda *a, **kw: None,
|
||||
config_picker=lambda args, keys: "default",
|
||||
config_picker=lambda args, keys: CaseKey.default(),
|
||||
)
|
||||
|
||||
def pattern(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
|
||||
|
||||
@@ -107,7 +107,7 @@ class TestPerTokenGroupFp8QuantConfigPicker:
|
||||
)
|
||||
|
||||
def test_config_picker_no_configs(self):
|
||||
config_keys: list[dict] = []
|
||||
config_keys: list[CaseKey] = []
|
||||
|
||||
args = _generate_fake_input(16, 4096, 128)
|
||||
selected_key = pick_config(args, config_keys)
|
||||
|
||||
@@ -397,6 +397,7 @@ class TestHelionKernelWrapper:
|
||||
)
|
||||
|
||||
assert wrapper._disabled is True
|
||||
assert wrapper._disabled_reason is not None
|
||||
assert "No configs available" in wrapper._disabled_reason
|
||||
|
||||
def test_disabled_wrapper_raises_on_call(self, sample_kernel):
|
||||
|
||||
@@ -89,7 +89,7 @@ class TestRmsNormDynamicPerTokenQuantConfigPicker:
|
||||
assert selected_key == CaseKey({"hidden_size": 2048, "num_tokens": 32})
|
||||
|
||||
def test_config_picker_no_configs(self):
|
||||
config_keys: list[dict] = []
|
||||
config_keys: list[CaseKey] = []
|
||||
|
||||
args = _generate_fake_input(16, 4096)
|
||||
selected_key = pick_config(args, config_keys)
|
||||
|
||||
@@ -112,7 +112,7 @@ class TestRmsNormPerBlockQuantConfigPicker:
|
||||
)
|
||||
|
||||
def test_config_picker_no_configs(self):
|
||||
config_keys: list[dict] = []
|
||||
config_keys: list[CaseKey] = []
|
||||
|
||||
args = _generate_fake_input(16, 4096, 128)
|
||||
selected_key = pick_config(args, config_keys)
|
||||
|
||||
@@ -96,7 +96,7 @@ class TestSiluAndMulPerBlockQuantConfigPicker:
|
||||
)
|
||||
|
||||
def test_config_picker_no_configs(self):
|
||||
config_keys: list[dict] = []
|
||||
config_keys: list[CaseKey] = []
|
||||
|
||||
args = _generate_fake_input(16, 4096, 128)
|
||||
selected_key = pick_config(args, config_keys)
|
||||
|
||||
@@ -83,7 +83,7 @@ class TestSiluMulFp8ConfigPicker:
|
||||
assert selected_key == CaseKey({"intermediate": 4096, "numtokens": 256})
|
||||
|
||||
def test_config_picker_no_configs(self):
|
||||
config_keys: list[dict] = []
|
||||
config_keys: list[CaseKey] = []
|
||||
|
||||
input_tensor = torch.randn(32, 4096, dtype=torch.bfloat16, device="cuda")
|
||||
scale = torch.tensor([0.5], dtype=torch.float32, device="cuda")
|
||||
@@ -355,6 +355,7 @@ class TestSiluMulFp8Integration:
|
||||
registered_kernels = get_registered_kernels()
|
||||
kernel_wrapper = registered_kernels["silu_mul_fp8"]
|
||||
fake_impl = kernel_wrapper._fake_impl
|
||||
assert fake_impl is not None
|
||||
|
||||
fake_output = fake_impl(input_tensor, scale)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user