[Bugfix][MiniMax-M3] Guard sparse decode sentinel indices (#54597)

Signed-off-by: Kyle Liang <kylliang@nvidia.com>
Signed-off-by: Kyle Liang <kyleliang-nv@users.noreply.github.com>
Co-authored-by: Kyle Liang <kyleliang-nv@users.noreply.github.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
kyleliang-nv
2026-10-08 18:27:57 +00:00
committed by GitHub
co-authored by Kyle Liang OpenAI Codex mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 972195f4f5
commit 1664c4f444
3 changed files with 128 additions and 10 deletions
+15 -2
View File
@@ -552,6 +552,21 @@ steps:
- pytest -v -s kernels/attention/test_deepgemm_attention.py
- pytest -v -s quantization/test_cutlass_w4a16.py
- label: ":nvidia: (B200) MiniMax-M3 Attention Kernels"
key: kernels-minimax-m3-attention-b200
timeout_in_minutes: 30
working_dir: "/vllm-workspace/"
device: b200-k8s
num_devices: 1
source_file_dependencies:
- vllm/models/minimax_m3/
- vllm/config/cache.py
- vllm/v1/kv_cache_interface.py
- tests/v1/attention/utils.py
- tests/kernels/attention/test_minimax_m3.py
commands:
- pytest -v -s tests/kernels/attention/test_minimax_m3.py
- label: ":nvidia: (B200) Kernels Shard %N"
key: kernels-b200
timeout_in_minutes: 30
@@ -601,7 +616,6 @@ steps:
- tests/kernels/test_fused_minimax_m3_qknorm_rope_kv_insert.py
- tests/kernels/mamba/test_gdn_prefill_cutedsl.py
- tests/kernels/test_bf16x3_router_gemm_cutedsl.py
- tests/kernels/attention/test_minimax_m3.py
- tests/kernels/test_ll_bf16_gemm.py
- tests/kernels/test_top_k_per_row.py
- tests/kernels/quantization/test_silu_mul_nvfp4_quant.py
@@ -647,7 +661,6 @@ steps:
pytest -v -s tests/kernels/moe/test_flashinfer.py
pytest -v -s tests/kernels/moe/test_trtllm_nvfp4_moe.py
pytest -v -s tests/kernels/mamba/test_gdn_prefill_cutedsl.py
pytest -v -s tests/kernels/attention/test_minimax_m3.py
;;
*)
echo "unexpected BUILDKITE_PARALLEL_JOB=$$BUILDKITE_PARALLEL_JOB" >&2
@@ -2612,6 +2612,103 @@ def test_decode_sparse_attention_correctness(
assert error.max().item() < 1.7e-2
@pytest.mark.parametrize("kv_layout", ["NHD", "HND"], indirect=True)
@pytest.mark.parametrize("capture_graph", [False, True])
@pytest.mark.parametrize("num_reqs", [2, 17], ids=["split-k", "single-chunk"])
@pytest.mark.parametrize(
"invalid_block", [-1, 1_000_000], ids=["sentinel", "out-of-range"]
)
def test_decode_sparse_attention_ignores_invalid_topk(
kv_layout: KVCacheLayout, capture_graph: bool, num_reqs: int, invalid_block: int
):
"""Unused top-k entries must not be dereferenced as block-table indices."""
torch.manual_seed(0)
decode_query_len = 4
seq_lens = torch.full(
(num_reqs,), decode_query_len, device="cuda", dtype=torch.int32
)
seq_lens[1] = 4096
q = torch.randn(
seq_lens.numel() * decode_query_len,
NUM_Q_HEADS,
HEAD_DIM,
device="cuda",
dtype=DTYPE,
)
kv_cache = _allocate_main_kv_via_contract(1, kv_layout)
# All logical blocks alias the single valid page. With a contiguous multi-row
# table, row 1's raw pointer offset -1 resolves to the final element of row
# 0. Poison that exact location so an unsafe load deterministically forms an
# out-of-range KV-cache pointer.
clean_block_table = torch.zeros(num_reqs, 32, device="cuda", dtype=torch.int32)
block_table = clean_block_table.clone()
block_table[0, -1] = 1_000_000
topk_idx = torch.full(
(NUM_KV_HEADS, q.shape[0], TOPK),
-1,
device="cuda",
dtype=torch.int32,
)
topk_idx[..., 0] = 0
# Request 1 has invalid entries both before and after its valid block.
# Two requests produce empty split-K chunks; 17 put the full top-k in one
# chunk, exercising neutral updates before and after a nonempty softmax.
request_topk = topk_idx[:, decode_query_len : 2 * decode_query_len]
request_topk.fill_(invalid_block)
request_topk[..., TOPK // 2] = 0
# A valid-only run isolates sentinel handling from BF16-vs-FP32 rounding.
# Include block 0 exactly once; replacing sentinels with 0 would duplicate
# its contribution. The general correctness test covers reference accuracy.
valid_topk = torch.zeros(
(NUM_KV_HEADS, q.shape[0], 1), device="cuda", dtype=torch.int32
)
expected = torch.empty_like(q)
minimax_m3_sparse_attn_decode(
q,
kv_cache,
valid_topk,
clean_block_table,
seq_lens,
NUM_KV_HEADS,
SM_SCALE,
expected,
decode_query_len,
)
actual = torch.empty_like(q)
def run_decode() -> None:
minimax_m3_sparse_attn_decode(
q,
kv_cache,
topk_idx,
block_table,
seq_lens,
NUM_KV_HEADS,
SM_SCALE,
actual,
decode_query_len,
)
if capture_graph:
# An unfixed kernel faults during capture and leaves the CUDA context
# unusable. Run negative-control reproductions in a dedicated process.
run_decode()
current_platform.synchronize()
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph):
run_decode()
for _ in range(3):
graph.replay()
else:
run_decode()
current_platform.synchronize()
assert torch.isfinite(actual).all()
torch.testing.assert_close(actual, expected, rtol=0, atol=0)
def test_decode_wrong_layout_breaks_parity():
"""Negative (AC-3/AC-5): consuming the physical HND buffer as if it were
already contiguous-NHD (i.e. skipping the allocator's inverse permute)
@@ -304,7 +304,7 @@ def _gqa_sparse_decode_kernel(
# attention range instead of letting padded rows produce negative lengths.
kv_len = tl.maximum(query_pos + 1, 0)
# Valid block count from seq_len (no sentinel): min(topk, cdiv(kv_len, blk)).
# Upper bound on top-k entries from causal length; entries may be sentinels.
idx_base = t_ptr + pid_kh * stride_th + pid_b * stride_tn
num_blocks = (kv_len + BLOCK_SIZE_K - 1) // BLOCK_SIZE_K
real_topk = tl.minimum(max_topk, num_blocks)
@@ -334,10 +334,15 @@ def _gqa_sparse_decode_kernel(
for _ in tl.range(chunk_start_topk, chunk_end_topk):
blk = tl.load(cur_idx_ptr).to(tl.int32)
cur_idx_ptr = cur_idx_ptr + stride_tk
c = blk * BLOCK_SIZE_K
page = tl.load(bt_row + blk).to(tl.int64)
# topk_idx uses -1 for unused entries. Select a safe logical block
# before dereferencing the block table; invalid and future blocks then
# contribute a neutral online-softmax update without control flow.
valid_blk = (blk >= 0) & (blk < num_blocks)
safe_blk = tl.where(valid_blk, blk, 0)
c = safe_blk * BLOCK_SIZE_K
page = tl.load(bt_row + safe_blk).to(tl.int64)
pos = c + off_n
pos_mask = pos < kv_len
pos_mask = (pos < kv_len) & valid_blk
k = tl.load(
kv_cache_ptr
+ page * stride_kv_blk
@@ -364,9 +369,11 @@ def _gqa_sparse_decode_kernel(
qk += tl.where(pos_mask[None, :], 0, float("-inf"))
qk += tl.dot(q, k) * sm_scale_log2e
m_ij = tl.maximum(m_i, tl.max(qk, axis=1))
p = tl.exp2(qk - m_ij[:, None])
m_safe = tl.where(valid_blk, m_ij, 0.0)
p = tl.where(pos_mask[None, :], tl.exp2(qk - m_safe[:, None]), 0.0)
l_ij = tl.sum(p, axis=1)
acc_o = acc_o * tl.exp2(m_i - m_ij)[:, None]
alpha = tl.where(valid_blk, tl.exp2(m_i - m_safe), 1.0)
acc_o = acc_o * alpha[:, None]
v = tl.load(
kv_cache_ptr
+ page * stride_kv_blk
@@ -390,8 +397,9 @@ def _gqa_sparse_decode_kernel(
)
v = (v * v_scale[:, None]).to(q.dtype)
acc_o += tl.dot(p.to(v.dtype), v)
m_i = m_ij
lse_i = m_ij + tl.log2(tl.exp2(lse_i - m_ij) + l_ij)
lse_ij = m_safe + tl.log2(tl.exp2(lse_i - m_safe) + l_ij)
m_i = tl.where(valid_blk, m_safe, m_i)
lse_i = tl.where(valid_blk, lse_ij, lse_i)
if USE_PDL:
tl.extra.cuda.gdc_launch_dependents()