[Perf][GLM] Fuse the kpool tail slot mapping into one Triton kernel (#57534)

Signed-off-by: Jared Wen <jaredwen@inferact.ai>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Jared Wen
2026-09-20 18:20:55 +08:00
committed by GitHub
co-authored by Claude Fable 5.1 mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent db1bfdd4fb
commit 1b9fa3eaa8
2 changed files with 120 additions and 2 deletions
@@ -391,3 +391,66 @@ def test_interleaved_decode_pollution_legacy_vs_circular():
# The circular mapping keeps the rings isolated under interleaving. # The circular mapping keeps the rings isolated under interleaving.
torch.testing.assert_close(circular, ground_truth) torch.testing.assert_close(circular, ground_truth)
@pytest.mark.skipif(not torch.cuda.is_available(), reason="needs a CUDA device")
@pytest.mark.parametrize(
"per_req,num_actual,padded_len",
[
([list(range(10)), list(range(12))], 22, 22),
([list(range(10)), list(range(12))], 22, 30),
([[3, 4], [0], [7, 8, 9]], 6, 8),
([[5]], 1, 1),
],
)
def test_triton_mapping_matches_cpu(per_req, num_actual, padded_len):
"""The CUDA (Triton) path must match the CPU torch reference, including
tokens between the last request boundary and num_actual_tokens (mapped to
the last request) and untouched padding beyond num_actual."""
positions, qsl, slot_mapping, _, num_reqs = make_batch(
per_req, padded_len=padded_len
)
# Replace the all--1 placeholder slots with sentinel values to check the
# padding range is copied through untouched.
slot_mapping = torch.arange(padded_len, dtype=torch.int64) + 1000
bt = make_tail_block_table(list(range(5, 5 + num_reqs)))
ref = circular_tail_slots(slot_mapping, bt, qsl, positions, num_actual, num_reqs)
got = circular_tail_slots(
slot_mapping.cuda(),
bt.cuda(),
qsl.cuda().to(torch.int32),
positions.cuda(),
num_actual,
num_reqs,
)
torch.testing.assert_close(got.cpu(), ref)
@pytest.mark.skipif(not torch.cuda.is_available(), reason="needs a CUDA device")
def test_triton_mapping_reads_strided_block_table():
"""The kernel must address the block table through its real stride(0),
not a dense assumption (the #57477 class of bug): a tail block table that
is a column view of a wider allocation must still read column 0."""
per_req = [list(range(10)), list(range(12))]
positions, qsl, slot_mapping, num_actual, num_reqs = make_batch(per_req)
own_blocks = [11, 22]
wide = torch.zeros(num_reqs, 8, dtype=torch.int32)
bt = wide[:, 1:4] # non-contiguous view, stride(0) == 8
bt[:, 0] = torch.tensor(own_blocks, dtype=torch.int32)
ref = circular_tail_slots(
slot_mapping, bt.contiguous(), qsl, positions, num_actual, num_reqs
)
got = circular_tail_slots(
slot_mapping.cuda(),
bt.cuda(),
qsl.cuda().to(torch.int32),
positions.cuda(),
num_actual,
num_reqs,
)
torch.testing.assert_close(got.cpu(), ref)
for req, blk in enumerate(own_blocks):
start, end = int(qsl[req]), int(qsl[req + 1])
assert (got[start:end] // KPOOL == blk).all()
+57 -2
View File
@@ -634,6 +634,38 @@ class DeepseekV32IndexerMetadata:
prefill: DeepseekV32IndexerPrefillMetadata | None = None prefill: DeepseekV32IndexerPrefillMetadata | None = None
@triton.jit(do_not_specialize=["num_reqs", "num_actual_tokens", "num_tokens"])
def _kpool_tail_slot_mapping_kernel(
slot_mapping_ptr,
block_table_ptr,
block_table_stride,
query_start_loc_ptr,
positions_ptr,
out_ptr,
num_reqs,
num_actual_tokens,
num_tokens,
kpool,
BLOCK: tl.constexpr,
):
pid = tl.program_id(0)
if pid < num_reqs:
start = tl.load(query_start_loc_ptr + pid)
# Tokens past the last request's boundary (if any) also map to it.
end = tl.load(query_start_loc_ptr + pid + 1)
end = tl.where(pid == num_reqs - 1, num_actual_tokens, end)
own_block = tl.load(block_table_ptr + pid * block_table_stride).to(tl.int64)
for i in range(start, end, BLOCK):
offs = i + tl.arange(0, BLOCK)
mask = offs < end
pos = tl.load(positions_ptr + offs, mask=mask, other=0).to(tl.int64)
tl.store(out_ptr + offs, own_block * kpool + pos % kpool, mask=mask)
else:
offs = num_actual_tokens + (pid - num_reqs) * BLOCK + tl.arange(0, BLOCK)
mask = offs < num_tokens
tl.store(out_ptr + offs, tl.load(slot_mapping_ptr + offs, mask=mask), mask=mask)
def compute_kpool_tail_slot_mapping( def compute_kpool_tail_slot_mapping(
slot_mapping: torch.Tensor, slot_mapping: torch.Tensor,
block_table: torch.Tensor, block_table: torch.Tensor,
@@ -646,10 +678,33 @@ def compute_kpool_tail_slot_mapping(
) -> torch.Tensor: ) -> torch.Tensor:
"""Map every token to its request's one circular tail block.""" """Map every token to its request's one circular tail block."""
if out is None: if out is None:
out = slot_mapping.clone() out = torch.empty_like(slot_mapping)
else: else:
assert out.shape == slot_mapping.shape assert out.shape == slot_mapping.shape
out.copy_(slot_mapping) if slot_mapping.is_cuda and slot_mapping.dim() == 1 and num_reqs > 0:
block = 256
num_tokens = slot_mapping.shape[0]
num_actual_tokens = min(num_actual_tokens, num_tokens)
grid = (num_reqs + triton.cdiv(num_tokens - num_actual_tokens, block),)
_kpool_tail_slot_mapping_kernel[grid](
slot_mapping,
block_table,
block_table.stride(0),
query_start_loc,
positions,
out,
num_reqs,
num_actual_tokens,
num_tokens,
kpool,
BLOCK=block,
num_warps=4,
)
return out
# Torch fallback: CPU tensors (the CPU unit tests), non-1D or empty inputs.
# Production always takes the Triton path above — spec-decode tokens arrive
# flattened token-major, so slot_mapping is 1D there too.
out.copy_(slot_mapping)
if num_actual_tokens == 0: if num_actual_tokens == 0:
return out return out
tokens = torch.arange(num_actual_tokens, device=slot_mapping.device) tokens = torch.arange(num_actual_tokens, device=slot_mapping.device)