biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Fix KV scales inconsistency in fp8 MLA & FlashInfer kv_cache_dtype "auto" leading to gibberish (#37054)

Browse Source 
Signed-off-by: Andy Lo <andy@mistral.ai>
This commit is contained in:
Andy Lo
2026-03-18 23:07:29 +00:00
committed by GitHub
parent 04244fd0e1
commit 577df69b26
8 changed files with 71 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
tests/v1/attention/test_mla_backends.py
View File

@@ -266,22 +266,6 @@ def create_and_prepopulate_kv_cache(
return kv_cache
class MockAttentionLayer:
"""A mock attention layer for testing."""
def __init__(self, device: torch.device):
self._q_scale = torch.tensor(1.0, device=device)
self._k_scale = torch.tensor(1.0, device=device)
self._v_scale = torch.tensor(1.0, device=device)
self._prob_scale = torch.tensor(1.0, device=device)
self._q_scale_float = 1.0
self._k_scale_float = 1.0
self._v_scale_float = 1.0
def forward(self, *_args, **_kwargs):
raise NotImplementedError
class MockSparseMLAAttentionLayer:
"""A mock sparse MLA attention layer for testing.
@@ -304,6 +288,8 @@ class MockSparseMLAAttentionLayer:
device: torch.device,
W_UK: torch.Tensor,
W_UV: torch.Tensor,
q_scale: float,
k_scale: float,
):
self.impl = impl
self.num_heads = num_heads
@@ -319,13 +305,13 @@ class MockSparseMLAAttentionLayer:
self.W_UV = W_UV.transpose(0, 1)
# Scale attributes needed by attention backends
self._q_scale = torch.tensor(1.0, device=device)
self._k_scale = torch.tensor(1.0, device=device)
self._v_scale = torch.tensor(1.0, device=device)
self._q_scale = torch.tensor(q_scale, device=device)
self._k_scale = torch.tensor(k_scale, device=device)
self._v_scale = torch.tensor(float("nan"), device=device)
self._prob_scale = torch.tensor(1.0, device=device)
self._q_scale_float = 1.0
self._k_scale_float = 1.0
self._v_scale_float = 1.0
self._q_scale_float = q_scale
self._k_scale_float = k_scale
self._v_scale_float = float("nan")
self._decode_concat_quant_fp8_op = _DecodeConcatQuantFP8(
static=True,
@@ -420,6 +406,8 @@ class MockMLAAttentionLayer(AttentionLayerBase):
kv_lora_rank: int,
device: torch.device,
kv_b_proj,
q_scale: float,
k_scale: float,
):
self.impl = impl
self.num_heads = num_heads
@@ -443,13 +431,13 @@ class MockMLAAttentionLayer(AttentionLayerBase):
self.W_UK_T = W_UK.permute(1, 2, 0)
# Scale attributes needed by attention backends
self._q_scale = torch.tensor(1.0, device=device)
self._k_scale = torch.tensor(1.0, device=device)
self._v_scale = torch.tensor(1.0, device=device)
self._q_scale = torch.tensor(q_scale, device=device)
self._k_scale = torch.tensor(k_scale, device=device)
self._v_scale = torch.tensor(float("nan"), device=device)
self._prob_scale = torch.tensor(1.0, device=device)
self._q_scale_float = 1.0
self._k_scale_float = 1.0
self._v_scale_float = 1.0
self._q_scale_float = q_scale
self._k_scale_float = k_scale
self._v_scale_float = float("nan")
self._decode_concat_quant_fp8_op = _DecodeConcatQuantFP8(
static=True,
@@ -568,6 +556,8 @@ def run_attention_backend(
qk_rope_head_dim: int,
v_head_dim: int,
mock_kv_b_proj,
q_scale: float,
k_scale: float,
kv_cache_dtype: str = "auto",
) -> torch.Tensor:
"""Run attention computation using the specified backend's AttentionImpl."""
@@ -625,6 +615,8 @@ def run_attention_backend(
kv_lora_rank=kv_lora_rank,
device=device,
kv_b_proj=mock_kv_b_proj,
q_scale=q_scale,
k_scale=k_scale,
)
# Populate static_forward_context with mock attention layers
@@ -674,6 +666,7 @@ def run_attention_backend(
@pytest.mark.parametrize("model", ["deepseek-ai/DeepSeek-R1"])
@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)])
def test_backend_correctness(
default_vllm_config,
dist_init,
@@ -681,6 +674,8 @@ def test_backend_correctness(
model: str,
tensor_parallel_size: int,
kv_cache_dtype: str,
q_scale: float,
k_scale: float,
):
"""
Test that all backends produce similar outputs to a reference implementation
@@ -709,6 +704,11 @@ def test_backend_correctness(
for b in BACKENDS_TO_TEST
if kv_cache_dtype in b.get_class().supported_kv_cache_dtypes
]
if (
q_scale != 1.0 or k_scale != 1.0
) and AttentionBackendEnum.CUTLASS_MLA in backends_to_test:
# CUTLASS_MLA does not support non-1 Q/K scales
backends_to_test.remove(AttentionBackendEnum.CUTLASS_MLA)
if not backends_to_test:
pytest.skip(f"No backends support kv_cache_dtype={kv_cache_dtype}")
@@ -1029,6 +1029,7 @@ def test_backend_correctness(
common_attn_metadata=common_attn_metadata,
randomize_blocks=True,
kv_cache_dtype=kv_cache_dtype,
scale=k_scale,
)
kv_cache_per_block_size[block_size] = kv_cache
@@ -1072,6 +1073,8 @@ def test_backend_correctness(
qk_rope_head_dim,
v_head_dim,
mock_kv_b_proj,
q_scale=q_scale,
k_scale=k_scale,
kv_cache_dtype=kv_cache_dtype,
)

11
tests/v1/attention/test_sparse_mla_backends.py
View File

@@ -178,6 +178,7 @@ def _quantize_dequantize_fp8_ds_mla(
@pytest.mark.parametrize("kv_cache_dtype", ["auto", "fp8", "fp8_ds_mla"])
@pytest.mark.parametrize("tensor_parallel_size", [1, 2, 4])
@pytest.mark.parametrize("block_size", [32, 64])
@pytest.mark.parametrize(("q_scale", "k_scale"), [(1.0, 1.0), (2.0, 3.0)])
def test_sparse_backend_decode_correctness(
default_vllm_config,
dist_init,
@@ -187,6 +188,8 @@ def test_sparse_backend_decode_correctness(
tensor_parallel_size,
block_size,
workspace_init,
q_scale: float,
k_scale: float,
):
if kv_cache_dtype not in backend_cls.supported_kv_cache_dtypes:
pytest.skip(f"{backend_cls.get_name()} does not support {kv_cache_dtype}")
@@ -332,7 +335,7 @@ def test_sparse_backend_decode_correctness(
kv_c_contexts, k_pe_contexts = [], []
reference_outputs = []
kv_cache_scale = torch.tensor(1.0, dtype=torch.float32, device=device)
kv_cache_scale = torch.tensor(k_scale, dtype=torch.float32, device=device)
global_token_idx = 0
for i in range(batch_spec.batch_size):
@@ -490,6 +493,8 @@ def test_sparse_backend_decode_correctness(
device=device,
W_UK=W_UK,
W_UV=W_UV,
q_scale=q_scale,
k_scale=k_scale,
)
out_buffer = torch.empty(
@@ -513,7 +518,9 @@ def test_sparse_backend_decode_correctness(
# FP8 quantization introduces some error, but should be within reasonable bounds
# BF16 (auto) should be very accurate, FP8 allows slightly more tolerance
if kv_cache_dtype.startswith("fp8"):
torch.testing.assert_close(backend_output, sdpa_reference, rtol=0.05, atol=0.05)
torch.testing.assert_close(
backend_output, sdpa_reference, rtol=0.065, atol=0.05
)
else:
torch.testing.assert_close(backend_output, sdpa_reference, rtol=0.01, atol=0.01)

12
tests/v1/attention/test_trtllm_attention_integration.py
View File

@@ -43,12 +43,12 @@ class MockAttentionLayer:
"""Minimal mock of an attention layer for testing."""
def __init__(self, device: torch.device):
self._q_scale = torch.tensor(1.0, device=device)
self._k_scale = torch.tensor(1.0, device=device)
self._v_scale = torch.tensor(1.0, device=device)
self._q_scale_float = 1.0
self._k_scale_float = 1.0
self._v_scale_float = 1.0
self._q_scale = torch.tensor(2.0, device=device)
self._k_scale = torch.tensor(3.0, device=device)
self._v_scale = torch.tensor(4.0, device=device)
self._q_scale_float = 2.0
self._k_scale_float = 3.0
self._v_scale_float = 4.0
self._o_scale_float = None