Enable DeepSeek V4 and GLM-5.1 on SM120 (#43477)
Signed-off-by: Zihua Wu <13583761+lucifer1004@users.noreply.github.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
co-authored by
Claude Opus 4.8
Yongye Zhu
parent
3ce15fd574
commit
44d95069e9
@@ -208,6 +208,7 @@ from vllm.config import (
|
||||
get_current_vllm_config,
|
||||
get_current_vllm_config_or_none,
|
||||
)
|
||||
from vllm.config.cache import CacheDType
|
||||
from vllm.distributed.parallel_state import (
|
||||
get_dcp_group,
|
||||
is_global_first_rank,
|
||||
@@ -319,6 +320,22 @@ def _detect_output_quant_key(
|
||||
return kFp8StaticTensorSym
|
||||
|
||||
|
||||
def _canonicalize_sparse_mla_kv_cache_dtype(
|
||||
attn_backend: type[AttentionBackend],
|
||||
kv_cache_dtype: CacheDType,
|
||||
) -> CacheDType:
|
||||
backend_name = attn_backend.get_name()
|
||||
if backend_name == "FLASHMLA_SPARSE" and is_quantized_kv_cache(kv_cache_dtype):
|
||||
return "fp8_ds_mla"
|
||||
if backend_name == "FLASHINFER_MLA_SPARSE_SM120" and kv_cache_dtype in (
|
||||
"auto",
|
||||
"fp8",
|
||||
"fp8_e4m3",
|
||||
):
|
||||
return "fp8_ds_mla"
|
||||
return kv_cache_dtype
|
||||
|
||||
|
||||
class MLAAttention(nn.Module, AttentionLayerBase):
|
||||
"""Multi-Head Latent Attention layer.
|
||||
|
||||
@@ -369,7 +386,7 @@ class MLAAttention(nn.Module, AttentionLayerBase):
|
||||
self.qk_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim
|
||||
|
||||
if cache_config is not None:
|
||||
kv_cache_dtype = cache_config.cache_dtype
|
||||
kv_cache_dtype: CacheDType = cache_config.cache_dtype
|
||||
calculate_kv_scales = cache_config.calculate_kv_scales
|
||||
else:
|
||||
kv_cache_dtype = "auto"
|
||||
@@ -393,24 +410,22 @@ class MLAAttention(nn.Module, AttentionLayerBase):
|
||||
num_heads=self.num_heads,
|
||||
)
|
||||
|
||||
# FlashMLA Sparse Attention fp8 backend uses "fp8_ds_mla" kv-cache format
|
||||
# Automatically convert fp8 kv-cache format to "fp8_ds_mla"
|
||||
if (
|
||||
self.attn_backend.get_name() == "FLASHMLA_SPARSE"
|
||||
and is_quantized_kv_cache(kv_cache_dtype)
|
||||
and kv_cache_dtype != "fp8_ds_mla"
|
||||
):
|
||||
assert cache_config is not None
|
||||
cache_config.cache_dtype = "fp8_ds_mla"
|
||||
kv_cache_dtype = "fp8_ds_mla"
|
||||
normalized_kv_cache_dtype = _canonicalize_sparse_mla_kv_cache_dtype(
|
||||
self.attn_backend, kv_cache_dtype
|
||||
)
|
||||
if normalized_kv_cache_dtype != kv_cache_dtype:
|
||||
if cache_config is not None:
|
||||
cache_config.cache_dtype = normalized_kv_cache_dtype
|
||||
kv_cache_dtype = normalized_kv_cache_dtype
|
||||
logger.info_once(
|
||||
"Using DeepSeek's fp8_ds_mla KV cache format. To use standard "
|
||||
"fp8 kv-cache format, please set `--attention-backend "
|
||||
"FLASHINFER_MLA_SPARSE`"
|
||||
"Using %s KV cache format for %s backend.",
|
||||
kv_cache_dtype,
|
||||
self.attn_backend.get_name(),
|
||||
)
|
||||
|
||||
if (
|
||||
self.attn_backend.get_name() == "FLASHINFER_MLA_SPARSE"
|
||||
and kv_cache_dtype != "fp8_ds_mla"
|
||||
and is_quantized_kv_cache(kv_cache_dtype)
|
||||
):
|
||||
logger.info_once(
|
||||
|
||||
@@ -23,24 +23,82 @@ def expert_num_tokens_round_up_and_sum(
|
||||
return torch.sum(ent).item()
|
||||
|
||||
|
||||
def compute_aligned_M_and_alignment(
|
||||
M: int,
|
||||
num_topk: int,
|
||||
local_num_experts: int,
|
||||
alignment: int,
|
||||
expert_tokens_meta: mk.ExpertTokensMetadata | None,
|
||||
) -> tuple[int, int]:
|
||||
"""Return (M_sum, alignment_used).
|
||||
|
||||
`alignment_used` may be smaller than the caller-supplied `alignment` on
|
||||
SM100/SM120 when DeepGEMM can JIT a smaller BLOCK_M for the per-call
|
||||
expected_m. Callers that index by block size (e.g. ``M_sum // block_m``)
|
||||
or assert workspace alignment must use the returned `alignment_used`,
|
||||
not their original `alignment` argument.
|
||||
|
||||
Prefer this over the int-returning :func:`compute_aligned_M` when the
|
||||
GEMM call site needs to wrap itself in ``mk_alignment_scope`` or
|
||||
otherwise reason about the actual per-expert padding.
|
||||
"""
|
||||
if (expert_tokens_meta is not None) and (
|
||||
expert_tokens_meta.expert_num_tokens_cpu is not None
|
||||
):
|
||||
return (
|
||||
expert_num_tokens_round_up_and_sum(
|
||||
expert_tokens_meta.expert_num_tokens_cpu, alignment=alignment
|
||||
),
|
||||
alignment,
|
||||
)
|
||||
|
||||
# expert_num_tokens not on cpu. Cap padding by min(M*num_topk,
|
||||
# local_num_experts) — at batch=1 decode only `num_topk` experts can be
|
||||
# active, so the worst-case `local_num_experts*(align-1)` is too loose.
|
||||
# Also shrink `alignment` to DeepGEMM's per-call theoretical BLOCK_M on
|
||||
# SM100/SM120 when smaller.
|
||||
expected_m = M * num_topk
|
||||
try:
|
||||
from vllm.utils.deep_gemm import (
|
||||
get_theoretical_mk_alignment_for_contiguous_layout,
|
||||
)
|
||||
|
||||
# num_groups=local_num_experts so the helper recovers per-expert em;
|
||||
# omitting it over-picks BLOCK_M on SM120 (heuristic assumes em is
|
||||
# already per-expert).
|
||||
per_call_align = get_theoretical_mk_alignment_for_contiguous_layout(
|
||||
expected_m=expected_m,
|
||||
num_groups=local_num_experts,
|
||||
)
|
||||
if per_call_align and per_call_align <= alignment:
|
||||
alignment = per_call_align
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
max_active_experts = min(M * num_topk, local_num_experts)
|
||||
M_sum = (M * num_topk) + max_active_experts * (alignment - 1)
|
||||
M_sum = round_up(M_sum, alignment)
|
||||
return M_sum, alignment
|
||||
|
||||
|
||||
def compute_aligned_M(
|
||||
M: int,
|
||||
num_topk: int,
|
||||
local_num_experts: int,
|
||||
alignment: int,
|
||||
expert_tokens_meta: mk.ExpertTokensMetadata | None,
|
||||
):
|
||||
if (expert_tokens_meta is not None) and (
|
||||
expert_tokens_meta.expert_num_tokens_cpu is not None
|
||||
):
|
||||
return expert_num_tokens_round_up_and_sum(
|
||||
expert_tokens_meta.expert_num_tokens_cpu, alignment=alignment
|
||||
)
|
||||
) -> int:
|
||||
"""Return ``M_sum`` only (backward-compat wrapper).
|
||||
|
||||
# expert_num_tokens information is not available on the cpu.
|
||||
# compute the max required size.
|
||||
M_sum = (M * num_topk) + local_num_experts * (alignment - 1)
|
||||
M_sum = round_up(M_sum, alignment)
|
||||
Equivalent to :func:`compute_aligned_M_and_alignment`'s first return
|
||||
value. Existing downstream callers and the warmup path that only size
|
||||
a workspace use this. Call sites that need the actual per-expert
|
||||
alignment (to wrap GEMMs in ``mk_alignment_scope``) should use
|
||||
:func:`compute_aligned_M_and_alignment` instead.
|
||||
"""
|
||||
M_sum, _ = compute_aligned_M_and_alignment(
|
||||
M, num_topk, local_num_experts, alignment, expert_tokens_meta
|
||||
)
|
||||
return M_sum
|
||||
|
||||
|
||||
@@ -51,12 +109,6 @@ def apply_expert_map(expert_id, expert_map):
|
||||
return expert_id
|
||||
|
||||
|
||||
@triton.jit
|
||||
def round_up_128(x: int) -> int:
|
||||
y = 128
|
||||
return ((x + y - 1) // y) * y
|
||||
|
||||
|
||||
@triton.jit
|
||||
def _fwd_kernel_ep_scatter_1(
|
||||
num_recv_tokens_per_expert,
|
||||
@@ -65,6 +117,7 @@ def _fwd_kernel_ep_scatter_1(
|
||||
num_experts: tl.constexpr,
|
||||
BLOCK_E: tl.constexpr,
|
||||
BLOCK_EXPERT_NUM: tl.constexpr,
|
||||
ALIGN_M: tl.constexpr,
|
||||
):
|
||||
cur_expert = tl.program_id(0)
|
||||
|
||||
@@ -74,7 +127,8 @@ def _fwd_kernel_ep_scatter_1(
|
||||
mask=offset_cumsum < num_experts,
|
||||
other=0,
|
||||
)
|
||||
tokens_per_expert = round_up_128(tokens_per_expert)
|
||||
# Round up to ALIGN_M so cumsum matches the workspace's per-expert slices.
|
||||
tokens_per_expert = ((tokens_per_expert + ALIGN_M - 1) // ALIGN_M) * ALIGN_M
|
||||
cumsum = tl.cumsum(tokens_per_expert) - tokens_per_expert
|
||||
|
||||
# Extract this block's offset from the register vector (warp shuffle,
|
||||
@@ -227,10 +281,12 @@ def ep_scatter(
|
||||
output_tensor_scale: torch.Tensor,
|
||||
m_indices: torch.Tensor,
|
||||
output_index: torch.Tensor,
|
||||
align_m: int = 128,
|
||||
block_size: int = 128,
|
||||
pack_ue8m0: bool = False,
|
||||
):
|
||||
BLOCK_E = 128 # token num of per expert is aligned to 128
|
||||
# BLOCK_E is the m_indices fill-loop tile (masked), independent of align_m.
|
||||
BLOCK_E = 128
|
||||
BLOCK_D = block_size # block size of activation-scale quantization
|
||||
num_warps = 8
|
||||
num_experts = num_recv_tokens_per_expert.shape[0]
|
||||
@@ -238,7 +294,7 @@ def ep_scatter(
|
||||
# grid = (triton.cdiv(hidden_size, BLOCK_D), num_experts)
|
||||
grid = num_experts
|
||||
|
||||
assert m_indices.shape[0] % BLOCK_E == 0
|
||||
assert m_indices.shape[0] % align_m == 0
|
||||
assert expert_start_loc.shape[0] == num_experts
|
||||
|
||||
# pack_ue8m0: scatter packs 4 UE8M0 bytes per int32; else copies scales as-is.
|
||||
@@ -253,6 +309,7 @@ def ep_scatter(
|
||||
num_warps=num_warps,
|
||||
BLOCK_E=BLOCK_E,
|
||||
BLOCK_EXPERT_NUM=triton.next_power_of_2(num_experts),
|
||||
ALIGN_M=align_m,
|
||||
)
|
||||
|
||||
grid = min(recv_topk.shape[0], 1024 * 8)
|
||||
@@ -418,7 +475,7 @@ def deepgemm_moe_permute(
|
||||
if block_size is not None:
|
||||
block_k = block_size
|
||||
|
||||
M_sum = compute_aligned_M(
|
||||
M_sum, align_used = compute_aligned_M_and_alignment(
|
||||
M=topk_ids.size(0),
|
||||
num_topk=topk_ids.size(1),
|
||||
local_num_experts=local_num_experts,
|
||||
@@ -482,11 +539,12 @@ def deepgemm_moe_permute(
|
||||
output_tensor_scale=aq_scale_out,
|
||||
m_indices=expert_ids,
|
||||
output_index=inv_perm,
|
||||
align_m=align_used,
|
||||
block_size=block_k,
|
||||
pack_ue8m0=pack_ue8m0,
|
||||
)
|
||||
|
||||
return aq_out, aq_scale_out, expert_ids, inv_perm
|
||||
return aq_out, aq_scale_out, expert_ids, inv_perm, align_used
|
||||
|
||||
|
||||
def deepgemm_unpermute_and_reduce(
|
||||
|
||||
@@ -318,11 +318,12 @@ class BatchedDeepGemmExperts(mk.FusedMoEExpertsModular):
|
||||
|
||||
def supports_packed_ue8m0_act_scales(self) -> bool:
|
||||
"""
|
||||
DeepGemm supports packed ue8m0 activation scales format in devices == sm100
|
||||
DeepGemm supports packed ue8m0 activation scales on Blackwell-family
|
||||
GPUs (SM100 datacenter and SM120 consumer).
|
||||
"""
|
||||
return (
|
||||
is_deep_gemm_e8m0_used()
|
||||
and current_platform.is_device_capability_family(100)
|
||||
return is_deep_gemm_e8m0_used() and (
|
||||
current_platform.is_device_capability_family(100)
|
||||
or current_platform.is_device_capability_family(120)
|
||||
)
|
||||
|
||||
def finalize_weight_and_reduce_impl(self) -> mk.TopKWeightAndReduce:
|
||||
|
||||
@@ -12,7 +12,7 @@ from vllm.model_executor.layers.fused_moe.config import (
|
||||
FusedMoEQuantConfig,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.deep_gemm_utils import (
|
||||
compute_aligned_M,
|
||||
compute_aligned_M_and_alignment,
|
||||
deepgemm_moe_permute,
|
||||
deepgemm_unpermute_and_reduce,
|
||||
)
|
||||
@@ -43,6 +43,7 @@ from vllm.utils.deep_gemm import (
|
||||
is_deep_gemm_supported,
|
||||
m_grouped_fp8_fp4_gemm_nt_contiguous,
|
||||
m_grouped_fp8_gemm_nt_contiguous,
|
||||
mk_alignment_scope,
|
||||
)
|
||||
from vllm.utils.import_utils import has_deep_gemm
|
||||
|
||||
@@ -210,10 +211,10 @@ class DeepGemmExperts(mk.FusedMoEExpertsModular):
|
||||
# Use the contiguous-layout M alignment (matches apply()); block_shape[0]
|
||||
# is the quant block (1 for MXFP8) and would under-size the workspace.
|
||||
block_m = get_mk_alignment_for_contiguous_layout()[0]
|
||||
M_sum = compute_aligned_M(
|
||||
M_sum, align_used = compute_aligned_M_and_alignment(
|
||||
M, topk, local_num_experts, block_m, expert_tokens_meta
|
||||
)
|
||||
assert M_sum % block_m == 0
|
||||
assert M_sum % align_used == 0
|
||||
|
||||
activation_out_dim = self.adjust_N_for_activation(N, activation)
|
||||
workspace1 = (M_sum, max(activation_out_dim, K))
|
||||
@@ -316,7 +317,7 @@ class DeepGemmExperts(mk.FusedMoEExpertsModular):
|
||||
|
||||
assert w2.size(1) == K
|
||||
|
||||
M_sum = compute_aligned_M(
|
||||
M_sum, _ = compute_aligned_M_and_alignment(
|
||||
M=topk_ids.size(0),
|
||||
num_topk=topk_ids.size(1),
|
||||
local_num_experts=local_num_experts,
|
||||
@@ -327,7 +328,7 @@ class DeepGemmExperts(mk.FusedMoEExpertsModular):
|
||||
a1q_perm = _resize_cache(
|
||||
workspace13.view(dtype=torch.float8_e4m3fn), (M_sum, K)
|
||||
)
|
||||
a1q, a1q_scale, expert_ids, inv_perm = deepgemm_moe_permute(
|
||||
a1q, a1q_scale, expert_ids, inv_perm, align_used = deepgemm_moe_permute(
|
||||
aq=a1q,
|
||||
aq_scale=a1q_scale,
|
||||
topk_ids=topk_ids,
|
||||
@@ -349,23 +350,35 @@ class DeepGemmExperts(mk.FusedMoEExpertsModular):
|
||||
else {}
|
||||
)
|
||||
|
||||
mm1_out = _resize_cache(workspace2, (M_sum, N))
|
||||
m_grouped_fp8_gemm_nt_contiguous(
|
||||
(a1q, a1q_scale), (w1, self.w1_scale), mm1_out, expert_ids, **gemm_kwargs
|
||||
)
|
||||
# Cap DG's BLOCK_M heuristic at the workspace's per-expert alignment;
|
||||
# otherwise the scheduler can pick the wrong expert id from m_indices
|
||||
# under cudagraph replay.
|
||||
with mk_alignment_scope(align_used):
|
||||
mm1_out = _resize_cache(workspace2, (M_sum, N))
|
||||
m_grouped_fp8_gemm_nt_contiguous(
|
||||
(a1q, a1q_scale),
|
||||
(w1, self.w1_scale),
|
||||
mm1_out,
|
||||
expert_ids,
|
||||
**gemm_kwargs,
|
||||
)
|
||||
|
||||
activation_out_dim = self.adjust_N_for_activation(N, activation)
|
||||
quant_out = _resize_cache(
|
||||
workspace13.view(dtype=torch.float8_e4m3fn), (M_sum, activation_out_dim)
|
||||
)
|
||||
a2q, a2q_scale = self._act_mul_quant(
|
||||
input=mm1_out.view(-1, N), output=quant_out, activation=activation
|
||||
)
|
||||
activation_out_dim = self.adjust_N_for_activation(N, activation)
|
||||
quant_out = _resize_cache(
|
||||
workspace13.view(dtype=torch.float8_e4m3fn), (M_sum, activation_out_dim)
|
||||
)
|
||||
a2q, a2q_scale = self._act_mul_quant(
|
||||
input=mm1_out.view(-1, N), output=quant_out, activation=activation
|
||||
)
|
||||
|
||||
mm2_out = _resize_cache(workspace2, (M_sum, K))
|
||||
m_grouped_fp8_gemm_nt_contiguous(
|
||||
(a2q, a2q_scale), (w2, self.w2_scale), mm2_out, expert_ids, **gemm_kwargs
|
||||
)
|
||||
mm2_out = _resize_cache(workspace2, (M_sum, K))
|
||||
m_grouped_fp8_gemm_nt_contiguous(
|
||||
(a2q, a2q_scale),
|
||||
(w2, self.w2_scale),
|
||||
mm2_out,
|
||||
expert_ids,
|
||||
**gemm_kwargs,
|
||||
)
|
||||
|
||||
if apply_router_weight_on_input:
|
||||
topk_weights = torch.ones_like(topk_weights)
|
||||
@@ -384,7 +397,8 @@ class DeepGemmFP4Experts(mk.FusedMoEExpertsModular):
|
||||
"""DeepGemm-based fused MoE expert implementation for FP4 weights.
|
||||
|
||||
Uses m_grouped_fp8_fp4_gemm_nt_contiguous with FP8 activations and
|
||||
MXFP4 (FP4 E2M1 packed as uint8) weights. Requires SM100+ (Blackwell).
|
||||
MXFP4 (FP4 E2M1 packed as uint8) weights. Requires Blackwell-family
|
||||
GPUs (SM100 datacenter or SM120 consumer).
|
||||
"""
|
||||
|
||||
# FP8 activation block size (hardcoded since mxfp4_w4a8 quant config
|
||||
@@ -409,9 +423,9 @@ class DeepGemmFP4Experts(mk.FusedMoEExpertsModular):
|
||||
def _supports_current_device() -> bool:
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
return (
|
||||
is_deep_gemm_supported()
|
||||
and current_platform.is_device_capability_family(100)
|
||||
return is_deep_gemm_supported() and (
|
||||
current_platform.is_device_capability_family(100)
|
||||
or current_platform.is_device_capability_family(120)
|
||||
)
|
||||
|
||||
@staticmethod
|
||||
@@ -454,10 +468,10 @@ class DeepGemmFP4Experts(mk.FusedMoEExpertsModular):
|
||||
activation: MoEActivation,
|
||||
) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
|
||||
block_m = get_mk_alignment_for_contiguous_layout()[0]
|
||||
M_sum = compute_aligned_M(
|
||||
M_sum, align_used = compute_aligned_M_and_alignment(
|
||||
M, topk, local_num_experts, block_m, expert_tokens_meta
|
||||
)
|
||||
assert M_sum % block_m == 0
|
||||
assert M_sum % align_used == 0
|
||||
|
||||
activation_out_dim = self.adjust_N_for_activation(N, activation)
|
||||
workspace1 = (M_sum, max(activation_out_dim, K))
|
||||
@@ -533,7 +547,7 @@ class DeepGemmFP4Experts(mk.FusedMoEExpertsModular):
|
||||
if global_num_experts == -1:
|
||||
global_num_experts = local_num_experts
|
||||
|
||||
M_sum = compute_aligned_M(
|
||||
M_sum, _ = compute_aligned_M_and_alignment(
|
||||
M=topk_ids.size(0),
|
||||
num_topk=topk_ids.size(1),
|
||||
local_num_experts=local_num_experts,
|
||||
@@ -544,7 +558,7 @@ class DeepGemmFP4Experts(mk.FusedMoEExpertsModular):
|
||||
a1q_perm = _resize_cache(
|
||||
workspace13.view(dtype=torch.float8_e4m3fn), (M_sum, K)
|
||||
)
|
||||
a1q, a1q_scale, expert_ids, inv_perm = deepgemm_moe_permute(
|
||||
a1q, a1q_scale, expert_ids, inv_perm, align_used = deepgemm_moe_permute(
|
||||
aq=a1q,
|
||||
aq_scale=a1q_scale,
|
||||
topk_ids=topk_ids,
|
||||
@@ -555,37 +569,40 @@ class DeepGemmFP4Experts(mk.FusedMoEExpertsModular):
|
||||
)
|
||||
assert a1q.size(0) == M_sum
|
||||
|
||||
# FC1: FP8 activations x FP4 weights
|
||||
# DeepGEMM 2.4.2 requires FP4-packed weights as int8 (kPackedFP4).
|
||||
mm1_out = _resize_cache(workspace2, (M_sum, N))
|
||||
m_grouped_fp8_fp4_gemm_nt_contiguous(
|
||||
(a1q, a1q_scale),
|
||||
(w1.view(torch.int8), self.w1_scale),
|
||||
mm1_out,
|
||||
expert_ids,
|
||||
recipe_a=(1, self._ACT_BLOCK_K),
|
||||
recipe_b=(1, self._WEIGHT_BLOCK_K),
|
||||
)
|
||||
# Cap DG's BLOCK_M heuristic at the workspace's per-expert alignment;
|
||||
# see DeepGemmExperts.apply for rationale.
|
||||
with mk_alignment_scope(align_used):
|
||||
# FC1: FP8 activations x FP4 weights
|
||||
# DeepGEMM 2.4.2 requires FP4-packed weights as int8 (kPackedFP4).
|
||||
mm1_out = _resize_cache(workspace2, (M_sum, N))
|
||||
m_grouped_fp8_fp4_gemm_nt_contiguous(
|
||||
(a1q, a1q_scale),
|
||||
(w1.view(torch.int8), self.w1_scale),
|
||||
mm1_out,
|
||||
expert_ids,
|
||||
recipe_a=(1, self._ACT_BLOCK_K),
|
||||
recipe_b=(1, self._WEIGHT_BLOCK_K),
|
||||
)
|
||||
|
||||
# SwiGLU activation + FP8 requant
|
||||
activation_out_dim = self.adjust_N_for_activation(N, activation)
|
||||
quant_out = _resize_cache(
|
||||
workspace13.view(dtype=torch.float8_e4m3fn), (M_sum, activation_out_dim)
|
||||
)
|
||||
a2q, a2q_scale = self._act_mul_quant(
|
||||
input=mm1_out.view(-1, N), output=quant_out, activation=activation
|
||||
)
|
||||
# SwiGLU activation + FP8 requant
|
||||
activation_out_dim = self.adjust_N_for_activation(N, activation)
|
||||
quant_out = _resize_cache(
|
||||
workspace13.view(dtype=torch.float8_e4m3fn), (M_sum, activation_out_dim)
|
||||
)
|
||||
a2q, a2q_scale = self._act_mul_quant(
|
||||
input=mm1_out.view(-1, N), output=quant_out, activation=activation
|
||||
)
|
||||
|
||||
# FC2: FP8 activations x FP4 weights
|
||||
mm2_out = _resize_cache(workspace2, (M_sum, K))
|
||||
m_grouped_fp8_fp4_gemm_nt_contiguous(
|
||||
(a2q, a2q_scale),
|
||||
(w2.view(torch.int8), self.w2_scale),
|
||||
mm2_out,
|
||||
expert_ids,
|
||||
recipe_a=(1, self._ACT_BLOCK_K),
|
||||
recipe_b=(1, self._WEIGHT_BLOCK_K),
|
||||
)
|
||||
# FC2: FP8 activations x FP4 weights
|
||||
mm2_out = _resize_cache(workspace2, (M_sum, K))
|
||||
m_grouped_fp8_fp4_gemm_nt_contiguous(
|
||||
(a2q, a2q_scale),
|
||||
(w2.view(torch.int8), self.w2_scale),
|
||||
mm2_out,
|
||||
expert_ids,
|
||||
recipe_a=(1, self._ACT_BLOCK_K),
|
||||
recipe_b=(1, self._WEIGHT_BLOCK_K),
|
||||
)
|
||||
|
||||
if apply_router_weight_on_input:
|
||||
topk_weights = torch.ones_like(topk_weights)
|
||||
|
||||
@@ -57,6 +57,40 @@ if has_triton_kernels():
|
||||
)
|
||||
|
||||
|
||||
def _pack_deepgemm_mxfp4_scales(
|
||||
w13_weight: torch.Tensor,
|
||||
w2_weight: torch.Tensor,
|
||||
w13_weight_scale: torch.Tensor,
|
||||
w2_weight_scale: torch.Tensor,
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
|
||||
deepgemm_post_process_weight_scale_block,
|
||||
)
|
||||
|
||||
num_experts = w13_weight.shape[0]
|
||||
intermediate_size_2 = w13_weight.shape[1] # = intermediate*2
|
||||
hidden_size = w13_weight.shape[2] * 2 # weight is FP4-packed
|
||||
intermediate_size = w2_weight.shape[2] * 2 # weight is FP4-packed
|
||||
block_shape = (1, 32) # MXFP4 block (per-row, K=32)
|
||||
|
||||
return (
|
||||
deepgemm_post_process_weight_scale_block(
|
||||
ws=w13_weight_scale.data,
|
||||
mn=intermediate_size_2,
|
||||
k=hidden_size,
|
||||
quant_block_shape=block_shape,
|
||||
num_groups=num_experts,
|
||||
),
|
||||
deepgemm_post_process_weight_scale_block(
|
||||
ws=w2_weight_scale.data,
|
||||
mn=hidden_size,
|
||||
k=intermediate_size,
|
||||
quant_block_shape=block_shape,
|
||||
num_groups=num_experts,
|
||||
),
|
||||
)
|
||||
|
||||
|
||||
class Mxfp4MoeBackend(Enum):
|
||||
NONE = "None"
|
||||
# DeepGEMM FP8xFP4 backend (SM100+)
|
||||
@@ -652,15 +686,18 @@ def convert_gpt_oss_weight_to_mxfp4_moe_kernel_format(
|
||||
"""Convert loaded weights into backend-specific kernel format."""
|
||||
|
||||
if mxfp4_backend == Mxfp4MoeBackend.DEEPGEMM_MXFP4:
|
||||
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
|
||||
_upcast_e8m0_to_fp32,
|
||||
w13_weight_scale, w2_weight_scale = _pack_deepgemm_mxfp4_scales(
|
||||
w13_weight,
|
||||
w2_weight,
|
||||
w13_weight_scale,
|
||||
w2_weight_scale,
|
||||
)
|
||||
|
||||
return (
|
||||
w13_weight.data,
|
||||
w2_weight.data,
|
||||
_upcast_e8m0_to_fp32(w13_weight_scale.data),
|
||||
_upcast_e8m0_to_fp32(w2_weight_scale.data),
|
||||
w13_weight_scale,
|
||||
w2_weight_scale,
|
||||
w13_bias,
|
||||
w2_bias,
|
||||
)
|
||||
@@ -1195,17 +1232,18 @@ def convert_weight_to_mxfp4_moe_kernel_format(
|
||||
"""
|
||||
|
||||
if mxfp4_backend == Mxfp4MoeBackend.DEEPGEMM_MXFP4:
|
||||
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
|
||||
_upcast_e8m0_to_fp32,
|
||||
w13_weight_scale, w2_weight_scale = _pack_deepgemm_mxfp4_scales(
|
||||
w13_weight,
|
||||
w2_weight,
|
||||
w13_weight_scale,
|
||||
w2_weight_scale,
|
||||
)
|
||||
|
||||
# Weights stay as uint8 packed FP4 — no layout change needed.
|
||||
# Convert E8M0 uint8 scales to float32.
|
||||
return (
|
||||
w13_weight.data,
|
||||
w2_weight.data,
|
||||
_upcast_e8m0_to_fp32(w13_weight_scale.data),
|
||||
_upcast_e8m0_to_fp32(w2_weight_scale.data),
|
||||
w13_weight_scale,
|
||||
w2_weight_scale,
|
||||
w13_bias,
|
||||
w2_bias,
|
||||
)
|
||||
|
||||
@@ -1058,6 +1058,34 @@ def _upcast_e8m0_to_fp32(scale: torch.Tensor) -> torch.Tensor:
|
||||
return fp32_bits.view(torch.float32)
|
||||
|
||||
|
||||
def deepgemm_post_process_weight_scale_block(
|
||||
ws: torch.Tensor,
|
||||
mn: int,
|
||||
k: int,
|
||||
quant_block_shape: tuple[int, ...],
|
||||
num_groups: int,
|
||||
is_sfa: bool = False,
|
||||
) -> torch.Tensor:
|
||||
if ws.dtype in (torch.float8_e8m0fnu, torch.uint8):
|
||||
# Scales already in E8M0 from checkpoint; upcast to fp32 and let
|
||||
# DeepGEMM pack the layout expected by the target architecture.
|
||||
ws = _upcast_e8m0_to_fp32(ws)
|
||||
else:
|
||||
assert ws.dtype == torch.float32, (
|
||||
f"Expected tensor scales dtype to be torch.float32 or "
|
||||
f"torch.float8_e8m0fnu or torch.uint8, got {ws.dtype} instead"
|
||||
)
|
||||
|
||||
return transform_sf_into_required_layout(
|
||||
sf=ws,
|
||||
mn=mn,
|
||||
k=k,
|
||||
recipe=(1, quant_block_shape[0], quant_block_shape[1]),
|
||||
num_groups=num_groups,
|
||||
is_sfa=is_sfa,
|
||||
)
|
||||
|
||||
|
||||
def deepgemm_post_process_fp8_weight_block(
|
||||
wq: torch.Tensor,
|
||||
ws: torch.Tensor,
|
||||
@@ -1073,13 +1101,13 @@ def deepgemm_post_process_fp8_weight_block(
|
||||
|
||||
if ws.dtype in (torch.float8_e8m0fnu, torch.uint8):
|
||||
# Scales already in E8M0 from checkpoint (float8_e8m0fnu, or raw E8M0
|
||||
# bits as uint8 for MXFP8) — upcast to fp32 and skip requantization
|
||||
# bits as uint8 for MXFP8) - upcast to fp32 and skip requantization
|
||||
# (weights already have power-of-two scales).
|
||||
ws = _upcast_e8m0_to_fp32(ws)
|
||||
else:
|
||||
assert ws.dtype == torch.float32, (
|
||||
f"Expected tensor scales dtype to be torch.float32 or "
|
||||
f"torch.float8_e8m0fnu, got {ws.dtype} instead"
|
||||
f"torch.float8_e8m0fnu or torch.uint8, got {ws.dtype} instead"
|
||||
)
|
||||
if use_e8m0:
|
||||
requant_weight_ue8m0_inplace(wq, ws, block_size=quant_block_shape)
|
||||
@@ -1094,16 +1122,12 @@ def deepgemm_post_process_fp8_weight_block(
|
||||
r = wq.size(0) // g
|
||||
wq = wq.view(g, r, d)
|
||||
ws = ws.view(g, r // quant_block_shape[0], d // quant_block_shape[1])
|
||||
# Pre-transform scale with recipe=(1, 128, 128) to broadcast + pack
|
||||
# into TMA-aligned UE8M0 (INT32) layout. At runtime fp8_einsum uses
|
||||
# recipe=(1, 1, 128) which sees INT dtype and skips re-transform.
|
||||
dg_ws = transform_sf_into_required_layout(
|
||||
sf=ws,
|
||||
dg_ws = deepgemm_post_process_weight_scale_block(
|
||||
ws=ws,
|
||||
mn=r,
|
||||
k=d,
|
||||
recipe=(1, quant_block_shape[0], quant_block_shape[1]),
|
||||
quant_block_shape=quant_block_shape,
|
||||
num_groups=g,
|
||||
is_sfa=False,
|
||||
)
|
||||
return wq, dg_ws
|
||||
|
||||
@@ -1113,22 +1137,12 @@ def deepgemm_post_process_fp8_weight_block(
|
||||
wq = wq.unsqueeze(0)
|
||||
ws = ws.unsqueeze(0)
|
||||
|
||||
# From https://github.com/deepseek-ai/DeepGEMM/blob/c9f8b34dcdacc20aa746b786f983492c51072870/csrc/utils/layout.hpp#L46
|
||||
# (1, block_n, block_k): (1, 128, 128) for FP8 block, (1, 1, 32) for MXFP8.
|
||||
recipe = (1, quant_block_shape[0], quant_block_shape[1])
|
||||
|
||||
# Ref : https://github.com/deepseek-ai/DeepGEMM/blob/c9f8b34dcdacc20aa746b786f983492c51072870/csrc/apis/gemm.hpp
|
||||
# DeepGemm uses the `transform_sf_into_required_layout` function to
|
||||
# represent scales in the correct format.
|
||||
dg_ws = transform_sf_into_required_layout(
|
||||
sf=ws,
|
||||
dg_ws = deepgemm_post_process_weight_scale_block(
|
||||
ws=ws,
|
||||
mn=wq.size(1),
|
||||
k=wq.size(2),
|
||||
recipe=recipe,
|
||||
quant_block_shape=quant_block_shape,
|
||||
num_groups=wq.size(0),
|
||||
# is the scale factors for A in (Refers to the argument A in A @ B).
|
||||
# Weights are B.
|
||||
is_sfa=False,
|
||||
)
|
||||
|
||||
if original_ndim == 2:
|
||||
|
||||
Reference in New Issue
Block a user