[Bugfix][ROCm] Fix MiniMax-M3 FP8 KV cache dtype (#45720)

Signed-off-by: Cam Quilici <cjquilici@gmail.com>
Signed-off-by: Cameron Quilici <cjquilici@gmail.com>
Co-authored-by: Hongxia Yang <62075498+hongxiayang@users.noreply.github.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
Cameron Quilici
2026-06-17 03:14:45 +00:00
committed by GitHub
co-authored by Hongxia Yang TJian
parent 556b063e45
commit efd15e192a
3 changed files with 60 additions and 5 deletions
@@ -15,11 +15,13 @@ from vllm.models.minimax_m3.common.ops.index_topk import (
minimax_m3_index_topk,
)
from vllm.models.minimax_m3.common.ops.sparse_attn import (
_FP8_DTYPES,
minimax_m3_sparse_attn,
minimax_m3_sparse_attn_decode,
)
from vllm.models.minimax_m3.common.sparse_attention import (
MiniMaxM3SparseBackend,
MiniMaxM3SparseTritonImpl,
)
from vllm.platforms import current_platform
from vllm.v1.attention.backends.utils import set_kv_cache_layout
@@ -79,6 +81,48 @@ SM_SCALE = HEAD_DIM**-0.5
TOPK = 16
@pytest.mark.parametrize(
("kv_cache_dtype", "expected_dtype"),
[
("fp8", current_platform.fp8_dtype()),
("fp8_e4m3", current_platform.fp8_dtype()),
(
"fp8_e5m2",
torch.float8_e5m2fnuz
if current_platform.is_fp8_fnuz()
else torch.float8_e5m2,
),
],
)
def test_sparse_impl_uses_platform_fp8_dtype(
kv_cache_dtype: str,
expected_dtype: torch.dtype,
):
impl = MiniMaxM3SparseTritonImpl(
num_heads=NUM_Q_HEADS,
head_size=HEAD_DIM,
scale=SM_SCALE,
num_kv_heads=NUM_KV_HEADS,
kv_cache_dtype=kv_cache_dtype,
topk_blocks=TOPK,
sparse_block_size=BLOCK_SIZE,
)
assert impl.kv_cache_fp8_dtype == expected_dtype
@pytest.mark.parametrize(
"dtype",
[
torch.float8_e4m3fn,
torch.float8_e4m3fnuz,
torch.float8_e5m2,
torch.float8_e5m2fnuz,
],
)
def test_sparse_kernels_recognize_fp8_dtypes(dtype: torch.dtype):
assert dtype in _FP8_DTYPES
# Index top-k kernels.
def _reference_index_topk(
idx_q: torch.Tensor,
@@ -24,6 +24,12 @@ from vllm.triton_utils import tl, triton
# One sparse block == one KV page.
SPARSE_BLOCK_SIZE = 128
_FP8_DTYPES = (
torch.float8_e4m3fn,
torch.float8_e4m3fnuz,
torch.float8_e5m2,
torch.float8_e5m2fnuz,
)
_SPARSE_ATTN_NUM_STAGES_KWARG: dict | None = None
@@ -456,7 +462,7 @@ def minimax_m3_sparse_attn(
batch = cu_seqlens_q.shape[0] - 1
topk = topk_idx.shape[-1]
gqa_group_size = num_heads // num_kv_heads
use_fp8 = kv_cache.dtype in (torch.float8_e4m3fn, torch.float8_e5m2)
use_fp8 = kv_cache.dtype in _FP8_DTYPES
grid = (max_query_len, num_kv_heads, batch)
_gqa_sparse_fwd_kernel[grid](
q,
@@ -513,7 +519,7 @@ def minimax_m3_sparse_attn_decode(
assert total_q == seq_lens.shape[0] * decode_query_len
max_topk = topk_idx.shape[-1]
gqa_group_size = num_heads // num_kv_heads
use_fp8 = kv_cache.dtype in (torch.float8_e4m3fn, torch.float8_e5m2)
use_fp8 = kv_cache.dtype in _FP8_DTYPES
use_pdl = current_platform.is_arch_support_pdl()
# `launch_pdl` is a Triton runtime kwarg only some backends accept (CUDA
# SM9+); this ROCm Triton rejects it even when False ("Keyword argument
@@ -291,9 +291,14 @@ class MiniMaxM3SparseImpl(AttentionImplBase[MiniMaxM3SparseMetadata]):
self.num_kv_heads = num_kv_heads if num_kv_heads is not None else num_heads
self.kv_cache_dtype = kv_cache_dtype
self.use_fp8_kv = is_quantized_kv_cache(kv_cache_dtype)
self.kv_cache_fp8_dtype = (
torch.float8_e5m2 if "e5m2" in kv_cache_dtype else torch.float8_e4m3fn
)
if "e5m2" in kv_cache_dtype:
self.kv_cache_fp8_dtype = (
torch.float8_e5m2fnuz
if current_platform.is_fp8_fnuz()
else torch.float8_e5m2
)
else:
self.kv_cache_fp8_dtype = current_platform.fp8_dtype()
# Sparse selection parameters (block_size == page size == SPARSE_BLOCK_SIZE).
self.topk_blocks = topk_blocks
self.block_size = sparse_block_size