[Kernel][DSv4.1] Bump FlashMLA to the 2026.09.30 release and its interleaved V4.1 KV cache (#61371)

Signed-off-by: juntianl <juntianl@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Kimi Code <noreply@moonshot.ai>
Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai>
This commit is contained in:
Juntian Liu
2026-10-15 13:58:25 -07:00
committed by GitHub
co-authored by Claude Opus 5.5 Kimi Code Yifan Qiao
parent 33212db2c2
commit f8460ef236
13 changed files with 262 additions and 165 deletions
+79 -20
View File
@@ -360,7 +360,7 @@ def _mxfp8_record_reference(
@pytest.mark.parametrize("compress_ratio", [1, 2])
def test_v41_rope_insert_mxfp8_record(compress_ratio: int):
"""The 528-byte V4.1 record quantizes the RoPE dims instead of keeping them
in bf16, so a page is [block_size x 512 fp8][block_size x 16 UE8M0].
in bf16, so a token is [512 fp8][16 UE8M0].
Byte-exact against the reference quantizer, including the rows and page
padding the kernel must leave alone.
@@ -398,9 +398,8 @@ def test_v41_rope_insert_mxfp8_record(compress_ratio: int):
values, scales = _mxfp8_record_reference(
latent[t], cos_sin[pos // compress_ratio * compress_ratio]
)
expected[page, row * 512 : (row + 1) * 512] = values
scale_offset = cache_block * 512 + row * 16
expected[page, scale_offset : scale_offset + 16] = scales
expected[page, row * 528 : row * 528 + 512] = values
expected[page, row * 528 + 512 : (row + 1) * 528] = scales
torch.testing.assert_close(cache_backing, expected, rtol=0, atol=0)
@@ -510,14 +509,13 @@ def test_v41_mxfp8_cache_round_trip():
values, scales = _ue8m0_reference(k[t].float(), 32, 448.0)
page, row = divmod(t, block_size)
torch.testing.assert_close(
cache[page, row * 512 : (row + 1) * 512],
cache[page, row * 528 : row * 528 + 512],
values.view(torch.uint8),
rtol=0,
atol=0,
)
scale_offset = block_size * 512 + row * 16
torch.testing.assert_close(
cache[page, scale_offset : scale_offset + 16],
cache[page, row * 528 + 512 : (row + 1) * 528],
(scales.log2() + 127).to(torch.uint8),
rtol=0,
atol=0,
@@ -564,7 +562,8 @@ def _decode_nvfp4_row(packed: torch.Tensor, scales: torch.Tensor) -> torch.Tenso
@pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only")
@pytest.mark.parametrize("compress_ratio", [1, 2])
def test_v41_rope_insert_nvfp4_record(compress_ratio: int):
"""The 288-byte V4.1 NVFP4 record: 256 B of e2m1 pairs then 32 e4m3 scales.
"""The 288-byte V4.1 NVFP4 record: 256 B of e2m1 pairs then 32 e4m3 scales
per token.
Scales are byte-exact against the reference (``amax / 6`` clamped to the
e4m3 range); values are checked as round-to-nearest on the e2m1 grid, whose
@@ -604,17 +603,13 @@ def test_v41_rope_insert_nvfp4_record(compress_ratio: int):
amax = rotated.view(32, 16).abs().amax(-1)
scale = (amax / 6.0).clamp(2.0**-9, 448.0).to(torch.float8_e4m3fn)
got_scales = cache_backing[page, cache_block * 256 + row * 32 :][:32]
record = cache_backing[page, row * 288 : (row + 1) * 288]
got_scales = record[256:]
torch.testing.assert_close(got_scales, scale.view(torch.uint8), rtol=0, atol=0)
decoded = _decode_nvfp4_row(
cache_backing[page, row * 256 : (row + 1) * 256], got_scales
)
decoded = _decode_nvfp4_row(record[:256], got_scales)
step = scale.float().repeat_interleave(16)
assert ((decoded - rotated).abs() <= step).all()
untouched[page, row * 256 : (row + 1) * 256] = cache_backing[
page, row * 256 : (row + 1) * 256
]
untouched[page, cache_block * 256 + row * 32 :][:32] = got_scales
untouched[page, row * 288 : (row + 1) * 288] = record
# Rows the kernel must not have touched, page padding included.
torch.testing.assert_close(cache_backing, untouched, rtol=0, atol=0)
@@ -670,10 +665,8 @@ def test_v41_nvfp4_gather_matches_insert(num_reqs, num_tokens, partial):
offset=offset,
)
expected = _decode_nvfp4_row(
backing[:, : block_size * 256].reshape(-1, 256),
backing[:, block_size * 256 : block_size * 288].reshape(-1, 32),
)
records = backing[:, : block_size * 288].reshape(-1, 288)
expected = _decode_nvfp4_row(records[:, :256], records[:, 256:])
def check(shorter_by):
assert (storage[:, num_tokens + offset :] == -123).all()
@@ -702,6 +695,72 @@ def test_v41_nvfp4_gather_matches_insert(num_reqs, num_tokens, partial):
check(1)
@pytest.mark.skipif(
not current_platform.is_device_capability_family(100),
reason="The V4.1 records are SM100-only",
)
@pytest.mark.parametrize("record", ["swa_fp8", "compressed_fp8", "compressed_fp4"])
def test_v41_cache_writers_match_flashmla_decode(record: str):
"""FlashMLA and vLLM's gathers read back the token a V4.1 writer put mid
page in a padded block, so all agree on the interleaved record layout."""
import vllm.v1.attention.ops.flashmla as fm
from vllm.models.deepseek_v41.common.ops.cache_utils import (
dequantize_and_gather_k_cache,
dequantize_and_gather_k_cache_triton,
)
from vllm.models.deepseek_v41.common.ops.fused_compress_quant_cache import (
rope_quant_insert,
)
device, block_size, topk = "cuda", 64, 64
slot = block_size + 37
bytes_per_token = 288 if record == "compressed_fp4" else 528
torch.manual_seed(0)
kv = torch.randn(1, 512, dtype=torch.bfloat16, device=device)
positions = torch.tensor([slot], device=device)
# cos = 1, sin = 0: RoPE is the identity, so the cache should hold `kv`.
cos_sin = torch.zeros(slot + 1, 64, device=device)
cos_sin[:, :32] = 1
# Padded like a packed block: a multiple of the record and of 512 B.
cache = torch.zeros(2, block_size + 32, bytes_per_token, dtype=torch.uint8)
cache = cache.to(device)[:, :block_size]
if record == "swa_fp8":
torch.ops._C.fused_deepseek_v4_kv_rope_insert(
kv, cache, positions, positions, cos_sin, block_size, None, True
)
else:
rope_quant_insert(kv, positions, cos_sin, cache, positions, 1)
indices = torch.full((1, 1, topk), -1, dtype=torch.int32, device=device)
indices[..., 0] = slot
extra = None
if record == "compressed_fp4":
cache, extra = torch.zeros(2, block_size, 528).to(cache), cache
out, _ = fm.flash_mla_with_kvcache(
q=torch.randn(1, 1, 64, 512, dtype=torch.bfloat16, device=device),
k_cache=cache.unsqueeze(2),
block_table=None,
cache_seqlens=None,
head_dim_v=512,
tile_scheduler_metadata=fm.get_mla_metadata()[0],
is_fp8_kvcache=True,
indices=indices if extra is None else torch.full_like(indices, -1),
extra_k_cache=None if extra is None else extra.unsqueeze(2),
extra_indices_in_kvcache=None if extra is None else indices,
)
for gather in (dequantize_and_gather_k_cache, dequantize_and_gather_k_cache_triton):
rows = torch.zeros(1, slot + 1, 512, dtype=torch.bfloat16, device=device)
block_table = torch.arange(2, dtype=torch.int32, device=device)[None]
seq_lens = torch.tensor([slot + 1], dtype=torch.int32, device=device)
k_cache = cache if extra is None else extra
gather(rows, k_cache, seq_lens, None, block_table, block_size, offset=0)
torch.testing.assert_close(
out[0, 0], rows[0, slot].expand(64, -1), rtol=0, atol=0
)
cos = torch.nn.functional.cosine_similarity(rows[0, slot].float(), kv[0].float(), 0)
assert cos > (0.98 if extra is not None else 0.999)
@pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only")
@pytest.mark.parametrize("compress_ratio", [1, 2])
@pytest.mark.parametrize("store_fp8", [False, True])