[Model][DeepSeek-V4.1] Permute mega-attention wq_b / wo_a while loading (#60064)

Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
aoshen02
2026-10-06 05:20:09 +00:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 533d81b4b0
commit 9a1f6fd4cb
5 changed files with 46 additions and 92 deletions
+25 -17
View File
@@ -3,25 +3,34 @@
"""Layout plumbing for DeepSeek V4.1 mega attention.
The kernel's Q and O layouts are produced by permuting ``wq_b`` rows and
``wo_a`` columns once at load, so what has to hold is that the permuted GEMMs
agree with a reference that permutes the activation instead.
``wo_a`` columns in their weight loaders, so what has to hold is that the
permuted GEMMs agree with a reference that permutes the activation instead.
"""
import torch
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.models.deepseek_v41.common.ops.fused_layout import (
WV_GROUP_SIZE,
o_fused_chunk_permutation,
o_fused_permutation,
permute_on_load,
permute_q_to_fused,
permute_wo_a_,
permute_wq_b_,
q_fused_permutation,
)
HEAD_DIM = 512
def _load(checkpoint: torch.Tensor, perm: torch.Tensor, dim: int) -> torch.Tensor:
"""Load ``checkpoint`` into a fresh parameter through a permuting loader."""
param = torch.nn.Parameter(torch.empty_like(checkpoint), requires_grad=False)
param.weight_loader = default_weight_loader # type: ignore[attr-defined]
permute_on_load(param, perm, dim)
param.weight_loader(param, checkpoint) # type: ignore[attr-defined]
return param.data
def test_permuted_wq_b_gemm_matches_permuted_activation():
"""Permuting wq_b's rows makes the Q GEMM emit the fused layout directly."""
torch.manual_seed(0)
@@ -32,18 +41,17 @@ def test_permuted_wq_b_gemm_matches_permuted_activation():
0, 256, (num_heads * HEAD_DIM, q_lora_rank // 32), dtype=torch.uint8
)
qr = torch.randn(num_tokens, q_lora_rank)
orig_weight, orig_scale = weight.clone(), scale.clone()
standard = (qr @ weight.T).view(num_tokens, num_heads, HEAD_DIM)
expected = permute_q_to_fused(standard)
permute_wq_b_(weight, scale, num_heads)
fused = (qr @ weight.T).view(num_tokens, num_heads, HEAD_DIM)
perm = q_fused_permutation(num_heads, HEAD_DIM)
loaded_weight = _load(weight, perm, dim=0)
fused = (qr @ loaded_weight.T).view(num_tokens, num_heads, HEAD_DIM)
torch.testing.assert_close(fused, expected)
# The scale has to follow its row, or dequant reads another row's scale.
perm = q_fused_permutation(num_heads, HEAD_DIM)
torch.testing.assert_close(weight, orig_weight[perm])
torch.testing.assert_close(scale, orig_scale[perm])
torch.testing.assert_close(loaded_weight, weight[perm])
torch.testing.assert_close(_load(scale, perm, dim=0), scale[perm])
def test_permuted_wo_a_consumes_fused_output():
@@ -54,18 +62,18 @@ def test_permuted_wo_a_consumes_fused_output():
weight = torch.randn(out_features, in_features)
scale = torch.randint(0, 256, (out_features, in_features // 32), dtype=torch.uint8)
standard_o = torch.randn(num_tokens, in_features)
orig_weight, orig_scale = weight.clone(), scale.clone()
expected = standard_o @ weight.T
# The kernel emits the same values in the fused chunk order.
perm = o_fused_permutation(WV_GROUP_SIZE, HEAD_DIM)
fused_o = standard_o[:, perm]
permute_wo_a_(weight, scale, WV_GROUP_SIZE)
torch.testing.assert_close(weight, orig_weight[:, perm])
torch.testing.assert_close(
scale, orig_scale[:, o_fused_chunk_permutation(WV_GROUP_SIZE, HEAD_DIM)]
)
loaded_weight = _load(weight, perm, dim=1)
torch.testing.assert_close(loaded_weight, weight[:, perm])
chunk_perm = o_fused_chunk_permutation(WV_GROUP_SIZE, HEAD_DIM)
torch.testing.assert_close(_load(scale, chunk_perm, dim=1), scale[:, chunk_perm])
# Unlike wq_b, this permutation sits inside the 4096-term reduction, so the
# summation order changes and the result differs in the last few ulps.
torch.testing.assert_close(fused_o @ weight.T, expected, rtol=1e-3, atol=1e-3)
torch.testing.assert_close(
fused_o @ loaded_weight.T, expected, rtol=1e-3, atol=1e-3
)
@@ -12,6 +12,8 @@ per-step shuffle.
import torch
from vllm.model_executor.model_loader.weight_utils import composed_weight_loader
HEAD_DIM = 512
Q_CHUNK = 16
O_CHUNK = 32
@@ -60,27 +62,13 @@ def _bytes_view(t: torch.Tensor) -> torch.Tensor:
return t.view(torch.uint8) if t.element_size() == 1 else t
def permute_wq_b_(
weight: torch.Tensor, weight_scale: torch.Tensor, num_local_heads: int
) -> None:
"""Permute an MXFP8 ``wq_b`` shard's rows and per-row scales, in place."""
head_dim = weight.shape[0] // num_local_heads
perm = q_fused_permutation(num_local_heads, head_dim).to(weight.device)
for t in (weight, weight_scale):
b = _bytes_view(t)
b.copy_(b[perm])
def permute_on_load(param: torch.Tensor, perm: torch.Tensor, dim: int) -> None:
"""Make ``param``'s weight loader gather the loaded shard by ``perm``."""
def gather(t: torch.Tensor) -> torch.Tensor:
return _bytes_view(t).index_select(dim, perm.to(t.device)).view(t.dtype)
def permute_wo_a_(
weight: torch.Tensor,
weight_scale: torch.Tensor,
heads_per_group: int = WV_GROUP_SIZE,
) -> None:
"""Permute an MXFP8 ``wo_a`` shard's input columns and scales, in place."""
head_dim = weight.shape[1] // heads_per_group
perm = o_fused_permutation(heads_per_group, head_dim).to(weight.device)
w = _bytes_view(weight)
w.copy_(w[:, perm])
chunk_perm = o_fused_chunk_permutation(heads_per_group, head_dim)
s = _bytes_view(weight_scale)
s.copy_(s[:, chunk_perm.to(weight.device)])
param.weight_loader = composed_weight_loader( # type: ignore[attr-defined]
param.weight_loader, # type: ignore[attr-defined]
gather,
)
-14
View File
@@ -500,22 +500,8 @@ class DSparkDeepseekV4ForCausalLM(nn.Module):
for layer in self.model.layers:
layer.ffn.finalize_mega_moe_weights()
def _finalize_attn(self) -> None:
"""Run the attention backend's post-load weight step on the draft layers.
They are ordinary DeepseekV4DecoderLayers, so they take the engine-wide
attention backend and owe it whatever the target model owes it -- mega
attention permutes wq_b / wo_a here and refuses to run without it.
Idempotent, like the target's.
"""
for layer in self.model.layers:
finalize = getattr(layer.attn, "finalize_loaded_weights", None)
if finalize is not None:
finalize()
def process_weights_after_loading(self) -> None:
self._finalize_moe()
self._finalize_attn()
def _remap_dspark_name(self, name: str) -> str | None:
"""Map a checkpoint ``mtp.{i}.*`` name to this model's parameter path.
@@ -32,8 +32,10 @@ from vllm.models.deepseek_v41.common.ops import (
)
from vllm.models.deepseek_v41.common.ops.fused_layout import (
WV_GROUP_SIZE,
permute_wo_a_,
permute_wq_b_,
o_fused_chunk_permutation,
o_fused_permutation,
permute_on_load,
q_fused_permutation,
)
from vllm.models.deepseek_v41.nvidia.flashmla import DeepseekV4FlashMLAAttention
from vllm.models.deepseek_v41.sparse_mla import (
@@ -182,7 +184,13 @@ class DeepseekV4MegaAttnAttention(DeepseekV4FlashMLAAttention):
"wo_a group."
)
self.n_wv_group = self.padded_heads // WV_GROUP_SIZE
self._fused_layouts_ready = False
q_perm = q_fused_permutation(self.n_local_heads, self.head_dim)
permute_on_load(self.wq_b.weight, q_perm, dim=0)
permute_on_load(self.wq_b.weight_scale, q_perm, dim=0)
o_perm = o_fused_permutation(WV_GROUP_SIZE, self.head_dim)
permute_on_load(self.wo_a.weight, o_perm, dim=1)
o_chunk_perm = o_fused_chunk_permutation(WV_GROUP_SIZE, self.head_dim)
permute_on_load(self.wo_a.weight_scale, o_chunk_perm, dim=1)
# ---- interface contract ------------------------------------------------
@@ -227,25 +235,6 @@ class DeepseekV4MegaAttnAttention(DeepseekV4FlashMLAAttention):
)
return self._wo_b_proj(z.flatten(1))
# ---- weights -----------------------------------------------------------
def finalize_loaded_weights(self) -> None:
"""Permute wq_b rows / wo_a columns into the kernel's layouts.
Idempotent: a second post-load pass must not permute twice.
"""
if self._fused_layouts_ready:
return
permute_wq_b_(
self.wq_b.weight.data, self.wq_b.weight_scale.data, self.n_local_heads
)
permute_wo_a_(
self.wo_a.weight.data,
self.wo_a.weight_scale.data,
self.n_local_heads // self.n_local_groups,
)
self._fused_layouts_ready = True
# ---- forward -----------------------------------------------------------
def forward_mqa(
@@ -255,11 +244,6 @@ class DeepseekV4MegaAttnAttention(DeepseekV4FlashMLAAttention):
positions: torch.Tensor,
output: QuantizedActivation,
) -> None:
if not self._fused_layouts_ready:
raise RuntimeError(
f"{self.prefix}: wq_b / wo_a were never permuted for the mega "
"attention kernel; refusing to run with mismatched layouts."
)
attn_metadata = get_forward_context().attn_metadata
if attn_metadata is None:
# Warmup dummy run: reserve the prefill workspace, produce zeros.
-12
View File
@@ -1364,17 +1364,6 @@ class DeepseekV4Model(nn.Module, EagleModelMixin):
for layer in islice(self.layers, self.start_layer, self.end_layer):
layer.ffn.finalize_mega_moe_weights()
def finalize_mega_attn_weights(self) -> None:
"""Permute wq_b / wo_a into FlashMLA's mega-attention layouts.
A no-op for every other attention layer, and idempotent, so a second
post-load pass cannot permute twice.
"""
for layer in islice(self.layers, self.start_layer, self.end_layer):
finalize = getattr(layer.attn, "finalize_loaded_weights", None)
if finalize is not None:
finalize()
def finalize_mhc_broadcast_weights(self) -> None:
if not get_pp_group().is_first_rank or self.start_layer >= self.end_layer:
return
@@ -1647,7 +1636,6 @@ class DeepseekV41LLMForCausalLM(
def process_weights_after_loading(self) -> None:
self.model.finalize_mega_moe_weights()
self.model.finalize_mhc_broadcast_weights()
self.model.finalize_mega_attn_weights()
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
return self.model.get_expert_mapping()