[Perf][DSV4.1] Optimize MegaMoE staging and NVFP4 cache gathers (#57604)

Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
Co-authored-by: Codex <noreply@openai.com>
This commit is contained in:
Woosuk Kwon
2026-09-18 12:57:19 -07:00
committed by GitHub
co-authored by Codex
parent 23e26e0588
commit 50d812b66a
5 changed files with 130 additions and 62 deletions
+68 -27
View File
@@ -473,15 +473,18 @@ _E2M1_MAGNITUDES = [0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0]
def _decode_nvfp4_row(packed: torch.Tensor, scales: torch.Tensor) -> torch.Tensor:
"""Unpack one 256-byte e2m1 row and apply its 32 e4m3 tile scales."""
"""Unpack e2m1 rows and apply their e4m3 tile scales."""
mags = torch.tensor(_E2M1_MAGNITUDES, device=packed.device)
codes = torch.empty(512, dtype=torch.uint8, device=packed.device)
codes[0::2] = packed & 0xF # even element in the low nibble
codes[1::2] = packed >> 4
codes = torch.empty(
(*packed.shape[:-1], 512), dtype=torch.uint8, device=packed.device
)
codes[..., 0::2] = packed & 0xF # even element in the low nibble
codes[..., 1::2] = packed >> 4
vals = mags[(codes & 7).long()] * torch.where(codes >= 8, -1.0, 1.0)
return (
vals.view(32, 16) * scales.view(torch.float8_e4m3fn).float().view(32, 1)
).flatten()
vals.reshape(*packed.shape[:-1], 32, 16)
* scales.view(torch.float8_e4m3fn).float().reshape(*packed.shape[:-1], 32, 1)
).flatten(-2)
@pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only")
@@ -543,7 +546,10 @@ def test_v41_rope_insert_nvfp4_record(compress_ratio: int):
@pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only")
def test_v41_nvfp4_gather_matches_insert():
@pytest.mark.parametrize("num_reqs", [1, 8, 32])
@pytest.mark.parametrize("num_tokens", [70, 128, 129, 2051])
@pytest.mark.parametrize("partial", [False, True])
def test_v41_nvfp4_gather_matches_insert(num_reqs, num_tokens, partial):
"""The NVFP4 gather dequantizes exactly what the insert kernel wrote."""
from vllm.models.deepseek_v41.common.ops import dequantize_and_gather_k_cache
from vllm.models.deepseek_v41.common.ops.fused_compress_quant_cache import (
@@ -553,38 +559,73 @@ def test_v41_nvfp4_gather_matches_insert():
torch.manual_seed(17)
device = "cuda"
block_size = 64
num_tokens = 70
num_blocks = 4
num_blocks = math.ceil(num_tokens / block_size)
page_bytes = math.ceil(block_size * 288 / 512) * 512
positions = torch.arange(num_tokens, dtype=torch.int64, device=device)
latent = torch.randn(num_tokens, 512, dtype=torch.bfloat16, device=device)
angles = torch.randn(128, 32, device=device)
angles = torch.randn(num_tokens, 32, device=device)
cos_sin = torch.cat((angles.cos(), angles.sin()), dim=-1)
backing = torch.zeros(num_blocks, page_bytes, dtype=torch.uint8, device=device)
cache = backing.as_strided((num_blocks, block_size, 288), (page_bytes, 288, 1))
slots = torch.arange(num_tokens, dtype=torch.int64, device=device)
rope_quant_insert(latent, positions, cos_sin, cache, slots, 1)
out = torch.zeros(1, num_tokens, 512, dtype=torch.bfloat16, device=device)
dequantize_and_gather_k_cache(
out,
cache,
seq_lens=torch.tensor([num_tokens], dtype=torch.int32, device=device),
gather_lens=None,
block_table=torch.arange(num_blocks, dtype=torch.int32, device=device).view(
1, -1
),
block_size=block_size,
offset=0,
offset = 7 if partial else 0
seq_lens = num_tokens - torch.arange(num_reqs, dtype=torch.int32, device=device)
gather_lens = seq_lens - 3 if partial else None
storage = torch.full(
(num_reqs, num_tokens + offset + 16, 512),
-123.0,
dtype=torch.bfloat16,
device=device,
)
for t in (0, 1, block_size, num_tokens - 1):
page, row = divmod(t, block_size)
expected = _decode_nvfp4_row(
backing[page, row * 256 : (row + 1) * 256],
backing[page, block_size * 256 + row * 32 :][:32],
out = storage[:, : num_tokens + offset]
block_table = torch.arange(num_blocks, dtype=torch.int32, device=device).repeat(
num_reqs, 1
)
def gather():
dequantize_and_gather_k_cache(
out,
cache,
seq_lens=seq_lens,
gather_lens=gather_lens,
block_table=block_table,
block_size=block_size,
offset=offset,
)
torch.testing.assert_close(out[0, t].float(), expected, rtol=0, atol=0)
expected = _decode_nvfp4_row(
backing[:, : block_size * 256].reshape(-1, 256),
backing[:, block_size * 256 : block_size * 288].reshape(-1, 32),
)
def check(shorter_by):
assert (storage[:, num_tokens + offset :] == -123).all()
for req in range(num_reqs):
start = 3 if partial else 0
length = num_tokens - req - start - shorter_by
assert (out[req, :offset] == -123).all()
assert (out[req, offset + length :] == -123).all()
torch.testing.assert_close(
out[req, offset : offset + length].float(),
expected[start : start + length],
rtol=0,
atol=0,
)
gather()
check(0)
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph):
gather()
storage.fill_(-123)
seq_lens.sub_(1)
if gather_lens is not None:
gather_lens.sub_(1)
graph.replay()
check(1)
@pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only")
+13 -9
View File
@@ -686,12 +686,14 @@ def test_deepseek_v4_mega_moe_does_not_double_add_fused_shared_expert(
not torch.cuda.is_available(),
reason="DeepSeek V4 MegaMoE fused input staging requires CUDA.",
)
def test_deepseek_v4_mega_moe_fused_input_staging_is_bitwise_exact():
@pytest.mark.parametrize("num_tokens", [7, 63, 64, 65, 127, 128, 135, 434, 16384])
@pytest.mark.parametrize("hidden_size", [256, 5120])
def test_deepseek_v4_mega_moe_fused_input_staging_is_bitwise_exact(
num_tokens, hidden_size
):
from vllm.third_party.deep_gemm.utils import per_token_cast_to_fp8
device = torch.device("cuda")
num_tokens = 7
hidden_size = 256
top_k = 8
generator = torch.Generator(device=device)
@@ -846,6 +848,7 @@ def test_deepseek_v4_mega_moe_stages_shared_scale_tma_layout(shared_block_m):
assert torch.equal(fused_x.view(torch.uint8), ref_x.view(torch.uint8))
assert torch.equal(fused_x_sf, ref_x_sf)
assert torch.equal(fused_shared_x_sf[populated], ref_shared_x_sf[populated])
assert torch.all(fused_shared_x_sf[~populated] == -1)
def test_deepseek_v4_pwal_hook_finalizes_mega_moe_and_mhc_broadcast():
@@ -884,11 +887,13 @@ def test_deepseek_v4_drafter_pwal_hooks_finalize_mega_moe():
reason="DeepSeek V4 MegaMoE fused input staging requires CUDA.",
)
@pytest.mark.parametrize("nonfinite_padding", [False, True])
def test_deepseek_v4_mega_moe_fused_input_staging_masks_padding(nonfinite_padding):
@pytest.mark.parametrize("num_tokens", [7, 65, 135])
def test_deepseek_v4_mega_moe_fused_input_staging_masks_padding(
nonfinite_padding, num_tokens
):
from vllm.third_party.deep_gemm.utils import per_token_cast_to_fp8
device = torch.device("cuda")
num_tokens = 7
hidden_size = 256
top_k = 8
@@ -916,10 +921,9 @@ def test_deepseek_v4_mega_moe_fused_input_staging_masks_padding(nonfinite_paddin
dtype=torch.float32,
generator=generator,
)
is_padding = torch.tensor(
[False, True, False, False, True, False, True],
device=device,
)
is_padding = torch.zeros(num_tokens, device=device, dtype=torch.bool)
is_padding[1::3] = True
is_padding[-1] = True
if nonfinite_padding:
topk_weights[is_padding] = float("nan")
@@ -12,7 +12,7 @@ import torch
from vllm.triton_utils import tl, triton
@triton.jit
@triton.jit(do_not_specialize=["num_tokens"])
def _prepare_megamoe_inputs_kernel(
hidden_states,
x_fp8,
@@ -40,27 +40,32 @@ def _prepare_megamoe_inputs_kernel(
topk_idx_stride_k: tl.constexpr,
topk_weights_out_stride_m: tl.constexpr,
topk_weights_out_stride_k: tl.constexpr,
num_tokens,
hidden_size: tl.constexpr,
top_k: tl.constexpr,
BLOCK_M: tl.constexpr,
BLOCK_K: tl.constexpr,
GROUP_K: tl.constexpr,
BLOCK_TOPK: tl.constexpr,
SHARED_BLOCK_M: tl.constexpr,
) -> None:
token_id = tl.program_id(0)
token_id = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
token_mask = token_id < num_tokens
k_block_id = tl.program_id(1)
k_offsets = k_block_id * BLOCK_K + tl.arange(0, BLOCK_K)
k_mask = k_offsets < hidden_size
k_mask = token_mask[:, None] & (k_offsets[None, :] < hidden_size)
hidden = tl.load(
hidden_states + token_id * hidden_stride_m + k_offsets * hidden_stride_k,
hidden_states
+ token_id[:, None] * hidden_stride_m
+ k_offsets[None, :] * hidden_stride_k,
mask=k_mask,
other=0.0,
).to(tl.float32)
num_groups: tl.constexpr = BLOCK_K // GROUP_K
hidden_groups = tl.reshape(tl.abs(hidden), [num_groups, GROUP_K])
amax = tl.max(hidden_groups, axis=1)
hidden_groups = tl.reshape(tl.abs(hidden), [BLOCK_M, num_groups, GROUP_K])
amax = tl.max(hidden_groups, axis=2)
amax = tl.maximum(amax, 1.0e-4)
scale = amax / 448.0
@@ -71,21 +76,24 @@ def _prepare_megamoe_inputs_kernel(
scale_exp = tl.minimum(tl.maximum(scale_exp, 1), 254)
rounded_scale = (scale_exp << 23).to(tl.float32, bitcast=True)
hidden_groups = tl.reshape(hidden, [num_groups, GROUP_K])
scaled = hidden_groups * (1.0 / rounded_scale)[:, None]
scaled = tl.reshape(scaled, [BLOCK_K])
hidden_groups = tl.reshape(hidden, [BLOCK_M, num_groups, GROUP_K])
scaled = hidden_groups * (1.0 / rounded_scale)[:, :, None]
scaled = tl.reshape(scaled, [BLOCK_M, BLOCK_K])
fp8 = scaled.to(tl.float8e4nv)
tl.store(
x_fp8 + token_id * x_stride_m + k_offsets * x_stride_k,
x_fp8 + token_id[:, None] * x_stride_m + k_offsets[None, :] * x_stride_k,
fp8,
mask=k_mask,
)
scale_offsets = tl.arange(0, num_groups)
packed_scale = tl.sum(scale_exp << (scale_offsets * 8), axis=0).to(tl.int32)
packed_scale = tl.sum(scale_exp << (scale_offsets[None, :] * 8), axis=1).to(
tl.int32
)
tl.store(
x_sf + token_id * x_sf_stride_m + k_block_id * x_sf_stride_k,
packed_scale,
mask=token_mask,
)
# DeepGEMM's SM100 shared-expert TMA loads require the activation scales
@@ -105,41 +113,48 @@ def _prepare_megamoe_inputs_kernel(
+ shared_row * shared_x_sf_stride_m
+ k_block_id * shared_x_sf_stride_k,
packed_scale,
mask=token_mask,
)
if k_block_id == 0:
topk_offsets = tl.arange(0, BLOCK_TOPK)
topk_mask = topk_offsets < top_k
token_is_padding = False
topk_mask = token_mask[:, None] & (topk_offsets[None, :] < top_k)
token_is_padding = tl.full((BLOCK_M,), False, tl.int1)
if is_padding is not None:
token_is_padding = tl.load(is_padding + token_id * is_padding_stride_m)
token_is_padding = tl.load(
is_padding + token_id * is_padding_stride_m,
mask=token_mask,
other=True,
)
ids = tl.load(
topk_ids + token_id * topk_ids_stride_m + topk_offsets * topk_ids_stride_k,
topk_ids
+ token_id[:, None] * topk_ids_stride_m
+ topk_offsets[None, :] * topk_ids_stride_k,
mask=topk_mask,
other=0,
).to(tl.int64)
ids = tl.where(token_is_padding, -1, ids)
ids = tl.where(token_is_padding[:, None], -1, ids)
tl.store(
topk_idx_out
+ token_id * topk_idx_stride_m
+ topk_offsets * topk_idx_stride_k,
+ token_id[:, None] * topk_idx_stride_m
+ topk_offsets[None, :] * topk_idx_stride_k,
ids,
mask=topk_mask,
)
weights = tl.load(
topk_weights
+ token_id * topk_weights_stride_m
+ topk_offsets * topk_weights_stride_k,
+ token_id[:, None] * topk_weights_stride_m
+ topk_offsets[None, :] * topk_weights_stride_k,
mask=topk_mask,
other=0.0,
)
weights = tl.where(token_is_padding, 0.0, weights)
weights = tl.where(token_is_padding[:, None], 0.0, weights)
tl.store(
topk_weights_out
+ token_id * topk_weights_out_stride_m
+ topk_offsets * topk_weights_out_stride_k,
+ token_id[:, None] * topk_weights_out_stride_m
+ topk_offsets[None, :] * topk_weights_out_stride_k,
weights,
mask=topk_mask,
)
@@ -195,7 +210,9 @@ def prepare_megamoe_inputs(
)
block_k = 128
grid = (num_tokens, triton.cdiv(hidden_size, block_k))
# On GB200, eight-row tiles win from 64 tokens; keep smaller batches untiled.
block_m = 8 if num_tokens >= 64 else 1
grid = (triton.cdiv(num_tokens, block_m), triton.cdiv(hidden_size, block_k))
block_topk = triton.next_power_of_2(top_k)
padding_stride_m = is_padding.stride(0) if is_padding is not None else 0
_prepare_megamoe_inputs_kernel[grid](
@@ -225,8 +242,10 @@ def prepare_megamoe_inputs(
topk_idx_out.stride(1),
topk_weights_out.stride(0),
topk_weights_out.stride(1),
num_tokens,
hidden_size,
top_k,
BLOCK_M=block_m,
BLOCK_K=block_k,
GROUP_K=32,
BLOCK_TOPK=block_topk,
@@ -584,7 +584,11 @@ def dequantize_and_gather_k_cache_triton(
num_reqs = seq_lens.shape[0]
NUM_WORKERS = 128
if k_cache.shape[-1] == V41_NVFP4_BYTES_PER_TOKEN:
_dequantize_and_gather_k_nvfp4_kernel[(num_reqs, NUM_WORKERS)](
# Keep short gathers unchanged; cap total workers for multi-request chunks.
num_workers = min(
max(128, 2048 // max(1, num_reqs)), max(128, out.shape[1] - offset)
)
_dequantize_and_gather_k_nvfp4_kernel[(num_reqs, num_workers)](
out,
out.stride(0),
out.stride(1),
@@ -422,7 +422,7 @@ class DeepseekV4MegaAttnAttention(DeepseekV4FlashMLAAttention):
if not swa_only:
assert flashmla_metadata is not None
dequantize_and_gather_k_cache(
kv_ws[:chunk_size],
kv_ws[:chunk_size, :chunk_n],
self._compressed_kv_cache(),
seq_lens=seq_lens[chunk_start:chunk_end] // self.compress_ratio,
gather_lens=None,