[3/N][warmup][DSv4] Migrate FA4 MLA and shared CuTeDSL kernels (#53565)
Signed-off-by: LopezCastroRoberto <rocastro@redhat.com> Co-authored-by: Codex <codex@openai.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
co-authored by
Codex
mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent
1713b9866a
commit
42801b3a6b
@@ -2,7 +2,7 @@
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Correctness test for the Inkling FA4 relative-attention score-mod kernel.
|
||||
|
||||
Checks ``INKLING_FA4_REL_ATTENTION_KERNEL`` against a pure-PyTorch reference
|
||||
Checks ``_INKLING_FA4_REL_ATTENTION_KERNEL`` against a pure-PyTorch reference
|
||||
that implements the relative bias exactly as documented in the Inkling architecture
|
||||
guide::
|
||||
|
||||
@@ -23,7 +23,7 @@ from vllm.models.inkling.nvidia.attention import (
|
||||
compute_log_scaling_tau,
|
||||
)
|
||||
from vllm.models.inkling.nvidia.ops.fa4_rel_attention import (
|
||||
INKLING_FA4_REL_ATTENTION_KERNEL,
|
||||
_INKLING_FA4_REL_ATTENTION_KERNEL,
|
||||
_use_sheared_bias,
|
||||
inkling_fa4_num_splits,
|
||||
)
|
||||
@@ -279,7 +279,7 @@ def _run_case(seq_lens, num_heads, num_kv_heads, rel_extent, window_left, seed=0
|
||||
num_kv_heads=num_kv_heads,
|
||||
max_kv_len=max(kv_lens),
|
||||
)
|
||||
out = INKLING_FA4_REL_ATTENTION_KERNEL(
|
||||
out = _INKLING_FA4_REL_ATTENTION_KERNEL(
|
||||
q,
|
||||
key_cache,
|
||||
value_cache,
|
||||
|
||||
@@ -10,6 +10,7 @@ Covers two things:
|
||||
+ standalone static-FP8-quant path it replaces (GPU-only, SM100/SM110).
|
||||
"""
|
||||
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import patch
|
||||
|
||||
import pytest
|
||||
@@ -99,6 +100,38 @@ def test_flash_attn_supports_quant_output_unknown_device():
|
||||
assert backend.supports_quant_output(kFp8StaticTensorSym) is False
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
("version", "enable_jit_warmup", "expected_calls"),
|
||||
[(4, True, 1), (4, False, 0), (3, True, 0)],
|
||||
)
|
||||
def test_flash_attn_registers_warmup_only_for_fa4(
|
||||
version: int,
|
||||
enable_jit_warmup: bool,
|
||||
expected_calls: int,
|
||||
):
|
||||
vllm_config = SimpleNamespace(
|
||||
kernel_config=SimpleNamespace(enable_jit_warmup=enable_jit_warmup)
|
||||
)
|
||||
with (
|
||||
patch(f"{_FA_MODULE}.flash_attn_varlen_func"),
|
||||
patch(f"{_FA_MODULE}.get_flash_attn_version", return_value=version),
|
||||
patch(f"{_FA_MODULE}._FA4_MLA_PREFILL_KERNEL.register_warmup") as register,
|
||||
patch(f"{_FA_MODULE}.current_platform") as platform,
|
||||
):
|
||||
platform.get_device_capability.return_value = DeviceCapability(10, 0)
|
||||
FlashAttnPrefillBackend(
|
||||
num_heads=16,
|
||||
scale=1.0,
|
||||
kv_lora_rank=512,
|
||||
qk_nope_head_dim=128,
|
||||
qk_rope_head_dim=64,
|
||||
v_head_dim=128,
|
||||
vllm_config=vllm_config,
|
||||
)
|
||||
|
||||
assert register.call_count == expected_calls
|
||||
|
||||
|
||||
def test_flash_attn_prefill_backend_signature_accepts_fused_kwargs():
|
||||
"""run_prefill_new_tokens must accept out/output_scale so the direct
|
||||
(non-**kwargs) call in forward_mha type- and runtime-checks."""
|
||||
|
||||
Reference in New Issue
Block a user