[Kernel] Enable fp8 support for pplx and BatchedTritonExperts. (#18864)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2025-07-03 14:55:40 -07:00
committed by GitHub
parent 2f2fcb31b8
commit 78fe77534b
25 changed files with 1277 additions and 663 deletions
@@ -184,15 +184,14 @@ class BatchedDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
def __init__(self,
max_num_tokens: int,
world_size: int,
dp_size: int,
num_dispatchers: int,
block_shape: list[int],
per_act_token_quant=False):
"""
max_num_tokens: Maximum number of tokens from a DP Rank
world_size: Number of EP ranks
dp_size: Number of data-parallel ranks
block_shape: Block quantization block shape
num_dispatchers: The number of DP dispatchers.
block_shape: Block quantization block shape.
per_act_token_quant: Per activation token quantization flag.
"""
super().__init__(
FusedMoEQuantConfig(
@@ -202,8 +201,7 @@ class BatchedDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
))
assert self.block_shape == self.DEEPGEMM_BLOCK_SHAPE
self.max_num_tokens = max_num_tokens
self.world_size = world_size
self.dp_size = dp_size
self.num_dispatchers = num_dispatchers
@property
def activation_formats(
@@ -233,7 +231,7 @@ class BatchedDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
# FIXME (varun): We should be able to dispatch only from the leader
# DP ranks in the case of TP > 1. At the moment, all the Ranks
# end up sending their tokens. This needs to be fixed.
num_dispatchers = self.world_size
num_dispatchers = self.num_dispatchers
num_experts = local_num_experts
max_num_tokens = a.size(
0) if self.max_num_tokens is None else self.max_num_tokens
@@ -15,8 +15,7 @@ class BatchedTritonOrDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
def __init__(self,
max_num_tokens: int,
world_size: int,
dp_size: int,
num_dispatchers: int,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
@@ -37,35 +36,28 @@ class BatchedTritonOrDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
block_shape=block_shape,
per_act_token_quant=per_act_token_quant,
))
self.max_num_tokens = max_num_tokens
self.world_size = world_size
self.dp_size = dp_size
self.allow_deep_gemm = allow_deep_gemm
# BatchedTritonKernel doesn't support block quantization
# at the moment.
self.batched_triton_experts = BatchedTritonExperts(
max_num_tokens=self.max_num_tokens,
world_size=self.world_size,
dp_size=self.dp_size,
max_num_tokens=max_num_tokens,
num_dispatchers=num_dispatchers,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=use_int8_w8a16,
use_int4_w4a16=use_int4_w4a16,
per_act_token_quant=self.per_act_token_quant,
block_shape=self.block_shape,
) if self.block_shape is None else None
)
is_fp8_128_block_quantized = (
use_fp8_w8a8 and self.block_shape
== BatchedDeepGemmExperts.DEEPGEMM_BLOCK_SHAPE)
self.allow_deep_gemm = (allow_deep_gemm and use_fp8_w8a8
and self.block_shape
== BatchedDeepGemmExperts.DEEPGEMM_BLOCK_SHAPE)
self.batched_deep_gemm_experts = BatchedDeepGemmExperts(
max_num_tokens=self.max_num_tokens,
world_size=self.world_size,
dp_size=self.dp_size,
max_num_tokens=max_num_tokens,
num_dispatchers=num_dispatchers,
block_shape=self.block_shape, # type: ignore[arg-type]
) if (self.allow_deep_gemm and is_fp8_128_block_quantized) else None
) if self.allow_deep_gemm else None
assert (self.batched_deep_gemm_experts is not None
or self.batched_triton_experts is not None)
@@ -138,12 +130,8 @@ class BatchedTritonOrDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
workspace2: torch.Tensor,
expert_num_tokens: Optional[torch.Tensor],
):
use_batched_deep_gemm_experts = (self.allow_deep_gemm
and self.batched_deep_gemm_experts
is not None)
experts = (self.batched_deep_gemm_experts
if use_batched_deep_gemm_experts else
self.batched_triton_experts)
if self.allow_deep_gemm else self.batched_triton_experts)
assert experts is not None
experts.apply(output, hidden_states, w1, w2, topk_ids, activation,
global_num_experts, expert_map, w1_scale, w2_scale,
+55 -9
View File
@@ -14,6 +14,7 @@ from vllm.distributed import get_dp_group, get_tensor_model_parallel_rank
from vllm.logger import init_logger
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig)
from vllm.utils import cdiv
logger = init_logger(__name__)
@@ -68,6 +69,57 @@ class FusedMoEQuantConfig:
# TODO: add col major flag?
# add detailed quant info for input, intermediates, weights, etc?
def __post_init__(self):
assert (not self.per_act_token_quant
or self.block_shape is None), "illegal quantization"
@property
def is_quantized(self) -> bool:
return self.quant_dtype is not None
@property
def is_per_act_token(self) -> bool:
return self.per_act_token_quant
@property
def is_block_quantized(self) -> bool:
return self.block_shape is not None
@property
def is_per_tensor(self) -> bool:
return not self.per_act_token_quant and self.block_shape is None
def scale_shape(
self,
max_tokens: int,
hidden_dim: int,
) -> Optional[tuple[int, int]]:
if self.is_quantized:
if self.is_block_quantized:
assert self.block_shape is not None
_, block_k = self.block_shape
k_tiles = cdiv(hidden_dim, block_k)
return (max_tokens, k_tiles)
elif self.is_per_act_token:
return (max_tokens, 1)
else:
return (1, 1)
else:
return None
def batched_scale_shape(
self,
num_experts: int,
max_tokens: int,
hidden_dim: int,
) -> Optional[tuple[int, int, int]]:
if self.is_quantized:
scale_shape = self.scale_shape(max_tokens, hidden_dim)
assert scale_shape is not None
return (num_experts, *scale_shape)
else:
return None
@staticmethod
def make(
use_fp8_w8a8: bool = False,
@@ -109,7 +161,6 @@ class FusedMoEParallelConfig:
tp_rank: int
dp_rank: int
ep_rank: int
world_size: int
use_ep: bool # whether to use EP or not
@@ -133,7 +184,7 @@ class FusedMoEParallelConfig:
and envs.VLLM_ALL2ALL_BACKEND == "deepep_low_latency")
@staticmethod
def make(tp_size_: int, dp_size_: int, world_size_: int,
def make(tp_size_: int, dp_size_: int,
vllm_parallel_config: ParallelConfig) -> "FusedMoEParallelConfig":
"""
Determine MoE parallel configuration. Based on the input tp_size_,
@@ -144,7 +195,6 @@ class FusedMoEParallelConfig:
tp_size_ (int): tp_size passed into the FusedMoE constructor.
dp_size_ (int): dp_size passed into the FusedMoE constructor.
ep_size_ (int): ep_size passed into the FusedMoE constructor.
world_size_ (int): the world size of the current All2All manager.
vllm_parallel_config (ParallelConfig): vllm's parallel config
object.
@@ -223,7 +273,6 @@ class FusedMoEParallelConfig:
dp_rank=dp_rank,
ep_size=1,
ep_rank=0,
world_size=world_size_,
use_ep=False)
# DP + EP / TP + EP / DP + TP + EP
assert use_ep
@@ -237,7 +286,6 @@ class FusedMoEParallelConfig:
dp_rank=dp_rank,
ep_size=ep_size,
ep_rank=ep_rank,
world_size=world_size_,
use_ep=True)
@@ -263,6 +311,8 @@ class FusedMoEConfig:
logger.debug("Using FusedMoEConfig::max_num_tokens=%d",
self.max_num_tokens)
assert self.max_num_tokens > 0
@property
def quant_dtype(self) -> Optional[torch.dtype]:
if self.quant_config is not None:
@@ -303,10 +353,6 @@ class FusedMoEConfig:
def ep_size(self):
return self.moe_parallel_config.ep_size
@property
def world_size(self):
return self.moe_parallel_config.world_size
@property
def tp_rank(self):
return self.moe_parallel_config.tp_rank
@@ -41,10 +41,7 @@ def run_cutlass_moe_fp8(
assert w2_scale is not None
assert w1.dtype == torch.float8_e4m3fn
assert w2.dtype == torch.float8_e4m3fn
if expert_num_tokens is None:
assert a1q.size(1) == w1.size(2), "Hidden size mismatch w1"
else:
assert a1q.size(2) == w1.size(2), "Hidden size mismatch w1"
assert a1q.size(-1) == w1.size(2), "Hidden size mismatch w1"
assert w1.size(1) == w2.size(2) * 2, "Hidden size mismatch w2"
assert w1_scale.dim() == 1 or w1_scale.size(
1) == 1 or w1_scale.shape[1] == w1.size(1), "W1 scale shape mismatch"
@@ -178,6 +175,8 @@ def run_cutlass_moe_fp8(
c2 = _resize_cache(workspace2, (M * topk, N))
c3 = _resize_cache(workspace13, (M * topk, K))
c1.fill_(0)
ops.cutlass_moe_mm(c1, a1q, w1, a1q_scale, w1_scale, expert_offsets,
problem_sizes1, ab_strides1, ab_strides1, c_strides1,
per_act_token, per_out_ch)
@@ -213,6 +212,7 @@ class CutlassExpertsFp8(mk.FusedMoEPermuteExpertsUnpermute):
per_act_token_quant: bool,
per_out_ch_quant: bool,
block_shape: Optional[list[int]] = None,
num_dispatchers: Optional[int] = None,
use_batched_format: bool = False,
):
super().__init__(
@@ -223,7 +223,9 @@ class CutlassExpertsFp8(mk.FusedMoEPermuteExpertsUnpermute):
block_shape=block_shape,
))
assert max_experts_per_worker > 0
assert not use_batched_format or num_dispatchers is not None
self.max_experts_per_worker = max_experts_per_worker
self.num_dispatchers = num_dispatchers
self.out_dtype = out_dtype
self.use_batched_format = use_batched_format
@@ -260,8 +262,12 @@ class CutlassExpertsFp8(mk.FusedMoEPermuteExpertsUnpermute):
output: tuple[int, ...] = ()
if self.use_batched_format:
padded_M = aq.size(1)
workspace1 = (self.max_experts_per_worker, padded_M, max(N, K))
workspace2 = (self.max_experts_per_worker, padded_M, (N // 2))
num_dp = self.num_dispatchers
assert num_dp is not None
workspace1 = (self.max_experts_per_worker, padded_M * num_dp,
max(N, K))
workspace2 = (self.max_experts_per_worker, padded_M * num_dp,
(N // 2))
output = (self.max_experts_per_worker, padded_M, K)
else:
workspace1 = (M * topk, max(2 * N, K))
@@ -16,12 +16,11 @@ class DeepEPHTPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
Prepare/Finalize using DeepEP High-Throughput kernels.
"""
def __init__(self, buffer: deep_ep.Buffer, world_size: int, rank: int,
def __init__(self, buffer: deep_ep.Buffer, num_dispatchers: int,
dp_size: int, rank_expert_offset: int):
super().__init__()
self.buffer = buffer
self.world_size = world_size
self.rank = rank
self.num_dispatchers_ = num_dispatchers
self.dp_size = dp_size
self.rank_expert_offset = rank_expert_offset
# The dispatch function returns a handle that the combine function
@@ -32,6 +31,9 @@ class DeepEPHTPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
# From https://github.com/deepseek-ai/DeepEP/blob/9fe9021f29c9083cd1808ab36b740208524d9f63/deep_ep/buffer.py#L164
self.available_rank_configs = [2, 4, 8, 16, 24, 32, 64, 128, 144, 160]
def num_dispatchers(self) -> int:
return self.num_dispatchers_
@property
def activation_format(self) -> mk.FusedMoEActivationFormat:
return mk.FusedMoEActivationFormat.Standard
@@ -136,20 +138,7 @@ class DeepEPHTPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
"apply_router_weight_on_input is only implemented for topk=1")
a1 = a1 * topk_weights.to(a1.dtype)
# Check if there is a block_shape / or if we can infer the quantization
# schemes from the scales.
per_token_quant = None
if all([
x is None
for x in [quant_config.block_shape, a1_scale, a2_scale]
]) and quant_config.quant_dtype is not None:
# Quantization required despite none of the inputs suggesting
# quantization. Fallback to per_token_dynamic quant.
per_token_quant = True
else:
per_token_quant = False
if per_token_quant:
if quant_config.per_act_token_quant:
a1q, a1q_scale = moe_kernel_quantize_input(
a1,
a1_scale,
@@ -7,7 +7,7 @@ import torch
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
from vllm.model_executor.layers.fused_moe.utils import (
maybe_fix_scales, moe_kernel_quantize_input)
moe_kernel_quantize_input, normalize_batched_scales_shape)
# DeepEP kernels quantize dispatch inputs in 128 element chunks.
DEEPEP_QUANT_BLOCK_SIZE = 128
@@ -42,20 +42,21 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
def __init__(self,
buffer: deep_ep.Buffer,
max_tokens_per_rank: int,
world_size: int,
dp_size: int,
num_dispatchers: int,
use_fp8_dispatch: bool = False):
super().__init__()
self.buffer = buffer
self.max_tokens_per_rank = max_tokens_per_rank
self.world_size = world_size
self.dp_size = dp_size
self.use_fp8_dispatch = use_fp8_dispatch
# The dispatch function returns a handle that the combine function
# requires. We store the handle here so it is available to the
# combine function.
self.handle = None
self.num_dispatchers_ = num_dispatchers
def num_dispatchers(self) -> int:
return self.num_dispatchers_
@property
def activation_format(self) -> mk.FusedMoEActivationFormat:
@@ -91,8 +92,6 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
assert isinstance(x, torch.Tensor)
assert not per_act_token_quant
num_experts, max_tokens, hidden_dim = x.size()
# TODO (varun): Optimization - Use a batched version of quant
@@ -104,7 +103,7 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
if quant_dtype is not None:
assert x_scales is not None
x_scales = maybe_fix_scales(x_scales, num_experts)
x_scales = normalize_batched_scales_shape(x_scales, num_experts)
return x, x_scales
@@ -12,42 +12,49 @@ from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
from vllm.model_executor.layers.fused_moe.fused_moe import (
get_config_dtype_str, try_get_optimal_moe_config)
from vllm.model_executor.layers.fused_moe.utils import (
_resize_cache, moe_kernel_quantize_input)
_resize_cache, moe_kernel_quantize_input, normalize_batched_scales_shape,
normalize_scales_shape)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
group_broadcast)
@triton.jit
def moe_mmk(
a_ptrs,
b_ptrs,
K,
expert_id,
a_scale_ptr,
b_scale_ptr,
# The stride variables represent how much to increase the ptr by when
# moving by 1 element in a particular dimension. E.g. `stride_am` is
# how much to increase `a_ptr` by to get the element one row down
# (A has M rows).
stride_ak,
stride_bk,
stride_asm,
stride_ask,
stride_bse,
stride_bsk,
stride_bsn,
# Offsets and masks
offs_m,
offs_n,
mask_m,
# Block size for block-wise quantization
group_n: tl.constexpr,
group_k: tl.constexpr,
# Meta-parameters
BLOCK_M: tl.constexpr,
BLOCK_N: tl.constexpr,
BLOCK_K: tl.constexpr,
compute_type: tl.constexpr,
use_w8a8: tl.constexpr,
use_w8a16: tl.constexpr):
a_ptrs,
b_ptrs,
K,
expert_id,
a_scale_ptr,
b_scale_ptr,
# The stride variables represent how much to increase the ptr by when
# moving by 1 element in a particular dimension. E.g. `stride_am` is
# how much to increase `a_ptr` by to get the element one row down
# (A has M rows).
stride_ak: tl.int64,
stride_bk: tl.int64,
stride_ase: tl.int64,
stride_asm: tl.int64,
stride_ask: tl.int64,
stride_bse: tl.int64,
stride_bsk: tl.int64,
stride_bsn: tl.int64,
# Offsets and masks
offs_m,
offs_n,
offs_bn,
mask_m,
# Block size for block-wise quantization
group_n: tl.constexpr,
group_k: tl.constexpr,
# Meta-parameters
BLOCK_M: tl.constexpr,
BLOCK_N: tl.constexpr,
BLOCK_K: tl.constexpr,
compute_type: tl.constexpr,
use_w8a8: tl.constexpr,
use_w8a16: tl.constexpr,
per_act_token_quant: tl.constexpr,
):
offs_k = tl.arange(0, BLOCK_K)
@@ -60,13 +67,22 @@ def moe_mmk(
# block-wise
if group_k > 0 and group_n > 0:
a_scale_ptrs = a_scale_ptr + offs_m * stride_asm
offs_bsn = offs_n // group_n
b_scale_ptrs = (b_scale_ptr + expert_id * stride_bse +
offs_bsn * stride_bsn)
offs_bsn = offs_bn // group_n
b_scale_ptrs = b_scale_ptr + offs_bsn * stride_bsn
# per act token
elif per_act_token_quant:
# Load per-token scale for activations
a_scale_ptrs = a_scale_ptr + offs_m * stride_asm
a_scale = tl.load(a_scale_ptrs, mask=mask_m, other=0.0)[:, None]
b_scale_ptrs = b_scale_ptr + offs_bn[None, :] * stride_bsn
b_scale = tl.load(b_scale_ptrs)
# tensor-wise
else:
a_scale = tl.load(a_scale_ptr)
b_scale = tl.load(b_scale_ptr + expert_id)
b_scale = tl.load(b_scale_ptr)
# -----------------------------------------------------------
# Iterate to compute a block of the C matrix.
@@ -96,13 +112,11 @@ def moe_mmk(
accumulator += tl.dot(a, b) * a_scale[:,
None] * b_scale[None, :]
else:
if use_w8a8:
# acc used to enable fp8_fast_accum
accumulator = tl.dot(a, b, acc=accumulator)
else:
accumulator += tl.dot(a, b)
# acc used to enable fp8_fast_accum
accumulator = tl.dot(a, b, acc=accumulator)
else:
accumulator += tl.dot(a, b)
# Advance the ptrs to the next K block.
a_ptrs += BLOCK_K * stride_ak
b_ptrs += BLOCK_K * stride_bk
@@ -122,47 +136,53 @@ def moe_mmk(
@triton.jit
def expert_triton_kernel(
a_ptr, #[max_tokens, K]
b_ptr, #[K, N]
c_ptr, #[max_tokens, N]
expert_id,
compute_type: tl.constexpr,
# Dimensions
M,
N,
K,
# Quantization data
a_scale_ptr,
b_scale_ptr,
b_zp_ptr,
# strides
stride_am,
stride_ak,
stride_bk,
stride_bn,
stride_cm,
stride_cn,
stride_asm,
stride_ask,
stride_bse,
stride_bsk,
stride_bsn,
# Blockwise quantization data
group_n,
group_k,
# Quantization schemes
use_fp8_w8a8: tl.constexpr,
use_int8_w8a16: tl.constexpr,
# Kernel config
BLOCK_M: tl.constexpr,
BLOCK_N: tl.constexpr,
BLOCK_K: tl.constexpr):
a_ptr, #[max_tokens, K]
b_ptr, #[K, N]
c_ptr, #[max_tokens, N]
expert_id,
compute_type: tl.constexpr,
# Dimensions
M,
N,
K,
# Quantization data
a_scale_ptr,
b_scale_ptr,
b_zp_ptr,
# strides
stride_am: tl.int64,
stride_ak: tl.int64,
stride_bk: tl.int64,
stride_bn: tl.int64,
stride_cm: tl.int64,
stride_cn: tl.int64,
stride_ase: tl.int64,
stride_asm: tl.int64,
stride_ask: tl.int64,
stride_bse: tl.int64,
stride_bsk: tl.int64,
stride_bsn: tl.int64,
# offsets
offs_bn,
# Blockwise quantization data
group_n,
group_k,
# Quantization schemes
use_fp8_w8a8: tl.constexpr,
use_int8_w8a16: tl.constexpr,
per_act_token_quant: tl.constexpr,
# Kernel config
BLOCK_M: tl.constexpr,
BLOCK_N: tl.constexpr,
BLOCK_K: tl.constexpr,
):
offs_m = tl.arange(0, BLOCK_M)
offs_n = tl.arange(0, BLOCK_N) % N
offs_k = tl.arange(0, BLOCK_K)
mask_m = offs_m < M
# Make grids of a + b pointers
a_ptrs = a_ptr + offs_m[:, None] * stride_am + offs_k[None, :] * stride_ak
b_ptrs = b_ptr + offs_k[:, None] * stride_bk + offs_n[None, :] * stride_bn
@@ -179,6 +199,7 @@ def expert_triton_kernel(
# (A has M rows).
stride_ak,
stride_bk,
stride_ase,
stride_asm,
stride_ask,
stride_bse,
@@ -187,6 +208,7 @@ def expert_triton_kernel(
# Offsets and masks
offs_m,
offs_n,
offs_bn,
mask_m,
# Block size for block-wise quantization
group_n,
@@ -197,7 +219,8 @@ def expert_triton_kernel(
BLOCK_K,
compute_type,
use_fp8_w8a8,
use_int8_w8a16)
use_int8_w8a16,
per_act_token_quant)
# store in C
offs_cn = tl.arange(0, BLOCK_N)
@@ -208,53 +231,57 @@ def expert_triton_kernel(
@triton.jit
def batched_triton_kernel(
a_ptr, # [E, max_num_tokens, K]
b_ptr, # [E, K, N]
c_ptr, # [E, max_num_tokens, N]
expert_num_tokens, # [E]
compute_type: tl.constexpr,
# Dimensions
max_num_tokens,
K,
N,
# Quantization data
a_scale_ptr,
b_scale_ptr,
b_zp_ptr,
# The stride variables represent how much to increase the ptr by when
# moving by 1 element in a particular dimension. E.g. `stride_am` is
# how much to increase `a_ptr` by to get the element one row down
# (A has M rows).
stride_ae,
stride_am,
stride_ak,
stride_be,
stride_bk,
stride_bn,
stride_ce,
stride_cm,
stride_cn,
stride_asm,
stride_ask,
stride_bse,
stride_bsk,
stride_bsn,
# Blockwise quantization data
group_n: tl.constexpr,
group_k: tl.constexpr,
# Quantization schemes
use_fp8_w8a8: tl.constexpr,
use_int8_w8a16: tl.constexpr,
# Kernel config
BLOCK_M: tl.constexpr,
BLOCK_N: tl.constexpr,
BLOCK_K: tl.constexpr):
a_ptr, # [E, max_num_tokens, K]
b_ptr, # [E, K, N]
c_ptr, # [E, max_num_tokens, N]
expert_num_tokens, # [E]
compute_type: tl.constexpr,
# Dimensions
max_num_tokens,
K,
N,
# Quantization data
a_scale_ptr,
b_scale_ptr,
b_zp_ptr,
# The stride variables represent how much to increase the ptr by when
# moving by 1 element in a particular dimension. E.g. `stride_am` is
# how much to increase `a_ptr` by to get the element one row down
# (A has M rows).
stride_ae: tl.int64,
stride_am: tl.int64,
stride_ak: tl.int64,
stride_be: tl.int64,
stride_bk: tl.int64,
stride_bn: tl.int64,
stride_ce: tl.int64,
stride_cm: tl.int64,
stride_cn: tl.int64,
stride_ase: tl.int64,
stride_asm: tl.int64,
stride_ask: tl.int64,
stride_bse: tl.int64,
stride_bsk: tl.int64,
stride_bsn: tl.int64,
# Blockwise quantization data
group_n: tl.constexpr,
group_k: tl.constexpr,
# Quantization schemes
use_fp8_w8a8: tl.constexpr,
use_int8_w8a16: tl.constexpr,
per_act_token_quant: tl.constexpr,
# Kernel config
BLOCK_M: tl.constexpr,
BLOCK_N: tl.constexpr,
BLOCK_K: tl.constexpr,
):
expert_id = tl.program_id(axis=0)
e_num_tokens = tl.load(expert_num_tokens + expert_id)
if e_num_tokens == 0:
# Early exit
return
# axis 1 is M_blocks * N_blocks
pid_mn = tl.program_id(axis=1)
#num_pid_m = tl.cdiv(max_num_tokens, BLOCK_M)
num_pid_n = tl.cdiv(N, BLOCK_N)
@@ -275,6 +302,16 @@ def batched_triton_kernel(
c_ptr = (c_ptr + expert_id * stride_ce + cta_m_start * stride_cm +
cta_n_start * stride_cn)
offs_bn = (pid_n * BLOCK_N + tl.arange(0, BLOCK_N).to(tl.int64)) % N
if use_fp8_w8a8:
a_scale_ptr = a_scale_ptr + expert_id * stride_ase
b_scale_ptr = b_scale_ptr + expert_id * stride_bse
# block-wise
if group_k > 0 and group_n > 0 or per_act_token_quant:
a_scale_ptr = a_scale_ptr + cta_m_start * stride_asm
expert_triton_kernel(
a_ptr,
b_ptr,
@@ -294,17 +331,21 @@ def batched_triton_kernel(
stride_bn,
stride_cm,
stride_cn,
stride_ase,
stride_asm,
stride_ask,
stride_bse,
stride_bsk,
stride_bsn,
# offsets
offs_bn,
# Blockwise quantization data
group_n,
group_k,
# Quantization schemes
use_fp8_w8a8,
use_int8_w8a16,
per_act_token_quant,
# Kernel config
BLOCK_M,
BLOCK_N,
@@ -326,6 +367,7 @@ def invoke_moe_batched_triton_kernel(
use_int8_w8a16: bool,
use_int4_w4a16: bool,
config: dict[str, int],
per_act_token_quant: bool,
block_shape: Optional[list[int]] = None):
assert not use_int4_w4a16
@@ -340,6 +382,42 @@ def invoke_moe_batched_triton_kernel(
grid = (expert_num_tokens.size(0), triton.cdiv(max_num_tokens, BLOCK_M) *
triton.cdiv(B.size(1), BLOCK_N))
A_scale = normalize_batched_scales_shape(A_scale,
expert_num_tokens.shape[0])
if B_scale is not None and B_scale.ndim == 1:
assert B_scale.numel() == expert_num_tokens.shape[0]
B_scale = B_scale.view(-1, 1, 1)
assert A_scale is None or A_scale.ndim == 3, (
f"{0 if A_scale is None else A_scale.shape}")
assert B_scale is None or B_scale.ndim == 1 or B_scale.ndim == 3, (
f"{0 if B_scale is None else B_scale.shape}")
if B_scale is not None:
if B_scale.ndim == 1:
stride_bse = 1
stride_bsk = 0
stride_bsn = 0
else:
stride_bse = B_scale.stride(0)
stride_bsk = B_scale.stride(2)
stride_bsn = B_scale.stride(1)
else:
stride_bse = 0
stride_bsk = 0
stride_bsn = 0
if A_scale is not None:
stride_ase = A_scale.stride(0)
stride_asm = A_scale.stride(1)
stride_ask = A_scale.stride(2)
else:
stride_ase = 0
stride_asm = 0
stride_ask = 0
batched_triton_kernel[grid](
A,
B,
@@ -364,17 +442,19 @@ def invoke_moe_batched_triton_kernel(
C.stride(0),
C.stride(1),
C.stride(2),
A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0,
A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0,
B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0,
B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0,
B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0,
stride_ase,
stride_asm,
stride_ask,
stride_bse,
stride_bsk,
stride_bsn,
# Blockwise quantization data
0 if block_shape is None else block_shape[0],
0 if block_shape is None else block_shape[1],
# Quantization schemes
use_fp8_w8a8,
use_int8_w8a16,
per_act_token_quant,
# Kernel config
BLOCK_M=BLOCK_M,
BLOCK_N=BLOCK_N,
@@ -391,15 +471,15 @@ class BatchedPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
def __init__(
self,
max_num_tokens: int,
world_size: int,
dp_size: int,
num_local_experts: int,
num_dispatchers: int,
rank: int,
):
super().__init__()
self.world_size = world_size
self.dp_size = dp_size
self.rank = rank
self.max_num_tokens = max_num_tokens
self.num_local_experts = num_local_experts
self.rank = rank
self.num_dispatchers_ = num_dispatchers
@property
def activation_format(self) -> mk.FusedMoEActivationFormat:
@@ -411,6 +491,9 @@ class BatchedPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
def topk_indices_dtype(self) -> Optional[torch.dtype]:
return None
def num_dispatchers(self) -> int:
return self.num_dispatchers_
def prepare(
self,
a1: torch.Tensor,
@@ -442,9 +525,7 @@ class BatchedPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
dtype=torch.int,
device=a1.device)
assert num_experts % self.world_size == 0
num_local_experts = num_experts // self.world_size
num_local_experts = self.num_local_experts
if quant_config.quant_dtype is None:
b_type = a1.dtype
@@ -456,21 +537,53 @@ class BatchedPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
dtype=b_type,
device=a1.device)
b_a1_scale = None
if quant_config.is_quantized:
scale_shape = quant_config.batched_scale_shape(
num_local_experts, self.max_num_tokens, hidden_dim)
assert quant_config.quant_dtype is None, "quantization NYI"
b_a1_scale = torch.empty(scale_shape,
dtype=torch.float32,
device=a1.device)
else:
assert a1_scale is None
b_a1_scale = None
first_expert = num_local_experts * self.rank
last_expert = first_expert + num_local_experts
a1_scale = normalize_scales_shape(a1_scale)
a2_scale = normalize_scales_shape(a2_scale)
for expert_id in range(first_expert, last_expert):
topks = torch.any(topk_ids == expert_id, dim=1).flatten()
rows = torch.count_nonzero(topks.flatten())
if rows == 0:
continue
idx = expert_id - first_expert
b_a1[idx, :rows, :] = a1[:topks.numel()][topks]
tokens_per_expert[idx] = rows
rhs = a1[:topks.numel()][topks]
if quant_config.quant_dtype is not None:
if a1_scale is not None:
if quant_config.is_per_act_token:
rhs_a1_scale = a1_scale[:topks.numel()][topks]
else:
rhs_a1_scale = a1_scale
else:
rhs_a1_scale = None
b_a1[idx, :rows, :], b_s = moe_kernel_quantize_input(
rhs,
rhs_a1_scale,
quant_config.quant_dtype,
quant_config.per_act_token_quant,
quant_config.block_shape,
)
assert b_s is not None
if quant_config.is_per_act_token:
b_a1_scale[idx, :rows] = b_s[:rows]
else:
b_a1_scale[idx, :b_s.shape[0]] = b_s
else:
b_a1[idx, :rows, :] = rhs
assert b_a1_scale is None or b_a1_scale.ndim == 3
@@ -514,8 +627,7 @@ class NaiveBatchedExperts(mk.FusedMoEPermuteExpertsUnpermute):
def __init__(
self,
max_num_tokens: int,
world_size: int,
dp_size: int,
num_dispatchers: int,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
@@ -532,13 +644,11 @@ class NaiveBatchedExperts(mk.FusedMoEPermuteExpertsUnpermute):
per_act_token_quant=per_act_token_quant,
block_shape=block_shape,
))
assert not use_fp8_w8a8, "NYI"
assert not use_int8_w8a8, "NYI"
assert not use_int8_w8a16, "NYI"
assert not use_int4_w4a16, "NYI"
self.max_num_tokens = max_num_tokens
self.world_size = world_size
self.dp_size = dp_size
self.num_dispatchers = num_dispatchers
@property
def activation_formats(
@@ -565,11 +675,21 @@ class NaiveBatchedExperts(mk.FusedMoEPermuteExpertsUnpermute):
local_num_experts: int,
) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...], torch.dtype]:
assert a.dim() == 2
num_dp = self.dp_size
num_dp = self.num_dispatchers
num_experts = local_num_experts
workspace13 = (num_experts, self.max_num_tokens * num_dp, K)
workspace2 = (self.max_num_tokens * num_dp, N)
return (workspace13, workspace2, workspace13, a.dtype)
output = workspace13
return (workspace13, workspace2, output, a.dtype)
def dequant(self, t: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
assert self.quant_config.is_quantized
f32 = torch.float32
if (self.quant_config.is_per_act_token
or self.quant_config.is_per_tensor):
return t.to(f32) * scale
else:
return t.to(f32) * group_broadcast(scale, t.shape)
def apply(
self,
@@ -612,9 +732,95 @@ class NaiveBatchedExperts(mk.FusedMoEPermuteExpertsUnpermute):
continue
tmp = _resize_cache(workspace2, (num, N))
input = hidden_states[expert, :num, :] @ w1[expert].transpose(0, 1)
self.activation(activation, tmp, input)
output[expert, :num, :] = tmp @ w2[expert].transpose(0, 1)
if self.quant_config.is_quantized:
assert a1q_scale is not None and w1_scale is not None
input = self.dequant(hidden_states[expert, :, :],
a1q_scale[expert])
w1_dq = self.dequant(w1[expert], w1_scale[expert])
input = input[:num] @ w1_dq.transpose(0, 1)
else:
input = hidden_states[expert, :num, :] @ w1[expert].transpose(
0, 1)
self.activation(activation, tmp, input.to(tmp.dtype))
if self.quant_config.is_quantized:
assert w2_scale is not None
w2_dq = self.dequant(w2[expert], w2_scale[expert])
else:
w2_dq = w2[expert]
output[expert, :num, :] = tmp @ w2_dq.transpose(0, 1).to(tmp.dtype)
def batched_moe_kernel_quantize_input(
A: torch.Tensor,
A_scale: Optional[torch.Tensor],
num_tokens: int,
E: int,
N: int,
expert_num_tokens: torch.Tensor,
qtype: Optional[torch.dtype],
per_act_token_quant: bool,
block_shape: Optional[list[int]] = None,
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
if (torch.compiler.is_compiling()
or torch.cuda.is_current_stream_capturing()):
# Note: this does a bunch of extra work because expert_num_tokens is
# ignored but it does support torch.compile + cudagraphs.
hidden_dim = A.size(-1)
assert A_scale is None or A_scale.ndim <= 2, (
f"{A_scale.shape if A_scale is not None else None}")
A_q, A_q_scale = moe_kernel_quantize_input(A.view(-1,
hidden_dim), A_scale,
qtype, per_act_token_quant,
block_shape)
A_q = A_q.view(E, -1, hidden_dim)
A_q_scale = normalize_batched_scales_shape(A_q_scale, E)
return A_q, A_q_scale
elif qtype is None:
return A, normalize_batched_scales_shape(A_scale, E)
else:
A_q = torch.empty_like(A, dtype=qtype)
if per_act_token_quant:
assert block_shape is None
scale_shape = (E, num_tokens, 1)
elif block_shape is not None:
_, block_k = block_shape
k_tiles = (A.shape[-1] + block_k - 1) // block_k
scale_shape = (E, num_tokens, k_tiles)
else:
scale_shape = (E, 1, 1)
A_q_scale = torch.zeros(scale_shape,
dtype=torch.float32,
device=A.device)
num_experts = expert_num_tokens.numel()
A_scale = normalize_batched_scales_shape(A_scale, num_experts)
for e in range(E):
num_tokens = int(expert_num_tokens[e].item())
if num_tokens > 0:
if A_scale is not None:
scales = A_scale[e, :min(num_tokens, A_scale.shape[1])]
else:
scales = None
A_q[e, :num_tokens], tmp_scale = moe_kernel_quantize_input(
A[e, :num_tokens],
scales,
qtype,
per_act_token_quant,
block_shape,
)
assert tmp_scale is not None
A_q_scale[e, :tmp_scale.shape[0]] = tmp_scale
return A_q, A_q_scale
class BatchedTritonExperts(mk.FusedMoEPermuteExpertsUnpermute):
@@ -627,8 +833,7 @@ class BatchedTritonExperts(mk.FusedMoEPermuteExpertsUnpermute):
def __init__(
self,
max_num_tokens: int,
world_size: int,
dp_size: int,
num_dispatchers: int,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
@@ -648,17 +853,14 @@ class BatchedTritonExperts(mk.FusedMoEPermuteExpertsUnpermute):
assert not use_int8_w8a8, "NYI"
assert not use_int8_w8a16, "NYI"
assert not use_int4_w4a16, "NYI"
assert max_num_tokens > 0
assert num_dispatchers > 0
self.use_fp8_w8a8 = use_fp8_w8a8
self.use_int8_w8a8 = use_int8_w8a8
self.use_int4_w4a16 = use_int4_w4a16
self.use_int8_w8a16 = use_int8_w8a16
self.max_num_tokens = max_num_tokens
self.world_size = world_size
self.dp_size = dp_size
assert world_size > 0
assert dp_size > 0
assert dp_size <= world_size
assert max_num_tokens > 0
self.num_dispatchers = num_dispatchers
@property
def activation_formats(
@@ -685,7 +887,7 @@ class BatchedTritonExperts(mk.FusedMoEPermuteExpertsUnpermute):
local_num_experts: int,
) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...], torch.dtype]:
assert a.dim() == 2
num_dp = self.world_size
num_dp = self.num_dispatchers
num_experts = local_num_experts
max_num_tokens = self.max_num_tokens
workspace13 = (num_experts, max_num_tokens * num_dp, max(K, N))
@@ -772,51 +974,48 @@ class BatchedTritonExperts(mk.FusedMoEPermuteExpertsUnpermute):
if self.use_fp8_w8a8:
intermediate_cache1.fill_(0)
a1q_scale = normalize_batched_scales_shape(a1q_scale, E)
# MM1
invoke_moe_batched_triton_kernel(A=hidden_states,
B=w1,
C=intermediate_cache1,
expert_num_tokens=expert_num_tokens,
compute_type=compute_type,
A_scale=a1q_scale,
B_scale=w1_scale,
B_zp=w1_zp,
use_fp8_w8a8=self.use_fp8_w8a8,
use_int8_w8a16=self.use_int8_w8a16,
use_int4_w4a16=self.use_int4_w4a16,
config=config,
block_shape=self.block_shape)
intermediate_cache2.fill_(0)
# TODO: would be nice to use expert_num_tokens here to reduce
# garbage compute
self.activation(activation, intermediate_cache2.view(-1, N // 2),
intermediate_cache1.view(-1, N))
ic2_hidden_size = intermediate_cache2.size(-1)
intermediate_cache2 = intermediate_cache2.view(-1, ic2_hidden_size)
qintermediate_cache2, a2q_scale = moe_kernel_quantize_input(
A=intermediate_cache2,
A_scale=a2_scale,
quant_dtype=self.quant_dtype,
invoke_moe_batched_triton_kernel(
A=hidden_states,
B=w1,
C=intermediate_cache1,
expert_num_tokens=expert_num_tokens,
compute_type=compute_type,
A_scale=a1q_scale,
B_scale=w1_scale,
B_zp=w1_zp,
use_fp8_w8a8=self.use_fp8_w8a8,
use_int8_w8a16=self.use_int8_w8a16,
use_int4_w4a16=self.use_int4_w4a16,
config=config,
per_act_token_quant=self.per_act_token_quant,
block_shape=self.block_shape)
qintermediate_cache2 = qintermediate_cache2.view(
(E, -1, ic2_hidden_size))
intermediate_cache2.fill_(0)
invoke_moe_batched_triton_kernel(A=qintermediate_cache2,
B=w2,
C=output,
expert_num_tokens=expert_num_tokens,
compute_type=compute_type,
A_scale=a2q_scale,
B_scale=w2_scale,
B_zp=w2_zp,
use_fp8_w8a8=self.use_fp8_w8a8,
use_int8_w8a16=self.use_int8_w8a16,
use_int4_w4a16=self.use_int4_w4a16,
config=config,
block_shape=self.block_shape)
# TODO (bnell): use triton utility from batched deep gemm.
self.activation(activation, intermediate_cache2.view(-1, N // 2),
intermediate_cache1.view(-1, N))
qintermediate_cache2, a2q_scale = batched_moe_kernel_quantize_input(
intermediate_cache2, a2_scale, max_num_tokens, E, N,
expert_num_tokens, self.quant_dtype, self.per_act_token_quant,
self.block_shape)
invoke_moe_batched_triton_kernel(
A=qintermediate_cache2,
B=w2,
C=output,
expert_num_tokens=expert_num_tokens,
compute_type=compute_type,
A_scale=a2q_scale,
B_scale=w2_scale,
B_zp=w2_zp,
use_fp8_w8a8=self.use_fp8_w8a8,
use_int8_w8a16=self.use_int8_w8a16,
use_int4_w4a16=self.use_int4_w4a16,
config=config,
per_act_token_quant=self.per_act_token_quant,
block_shape=self.block_shape)
@@ -1127,6 +1127,8 @@ def dispatch_fused_experts_func(inplace: bool) -> Callable[..., torch.Tensor]:
return torch_vllm_outplace_fused_experts
# TODO (bnell): replace this with modular op. Can get rid of inplace/outplace
# torch ops.
def fused_experts(hidden_states: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
+12 -22
View File
@@ -14,7 +14,6 @@ import vllm.envs as envs
from vllm.config import get_current_vllm_config
from vllm.distributed import (get_dp_group, get_ep_group,
get_tensor_model_parallel_world_size,
get_world_group,
tensor_model_parallel_all_reduce)
from vllm.distributed.eplb.eplb_state import EplbState
from vllm.forward_context import ForwardContext, get_forward_context
@@ -114,6 +113,9 @@ class FusedMoEMethodBase(QuantizeMethodBase):
hidden_dim_scale_bytes=hidden_scale_bytes,
)
num_dispatchers = (all2all_manager.world_size //
all2all_manager.tp_group.world_size)
# Intranode pplx a2a takes a group name while internode does not.
if not all2all_manager.internode:
all_to_all_args[
@@ -124,10 +126,8 @@ class FusedMoEMethodBase(QuantizeMethodBase):
prepare_finalize = PplxPrepareAndFinalize(
handle,
max_num_tokens=moe.max_num_tokens,
world_size=all2all_manager.world_size,
rank=all2all_manager.rank,
# dp_size actually means tp_size, bug in pplx kernels
dp_size=all2all_manager.tp_group.world_size,
num_local_experts=moe.num_local_experts,
num_dispatchers=num_dispatchers,
)
elif moe.use_deepep_ht_kernels:
assert moe.dp_size == all2all_manager.dp_world_size
@@ -136,16 +136,13 @@ class FusedMoEMethodBase(QuantizeMethodBase):
handle = all2all_manager.get_handle(all_to_all_args)
prepare_finalize = DeepEPHTPrepareAndFinalize(
handle,
world_size=all2all_manager.world_size,
rank=all2all_manager.rank,
num_dispatchers=all2all_manager.world_size,
dp_size=all2all_manager.dp_world_size,
rank_expert_offset=all2all_manager.rank *
moe.num_local_experts,
)
elif moe.use_deepep_ll_kernels:
assert moe.dp_size == all2all_manager.dp_world_size
all_to_all_args = dict(
max_num_tokens_per_dp_rank=moe.max_num_tokens,
token_hidden_size=moe.hidden_dim,
@@ -168,8 +165,7 @@ class FusedMoEMethodBase(QuantizeMethodBase):
prepare_finalize = DeepEPLLPrepareAndFinalize(
handle,
max_tokens_per_rank=moe.max_num_tokens,
world_size=all2all_manager.world_size,
dp_size=all2all_manager.dp_world_size,
num_dispatchers=all2all_manager.world_size,
use_fp8_dispatch=use_fp8_dispatch,
)
@@ -245,18 +241,12 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
assert self.fused_experts == fused_experts
all2all_manager = get_ep_group().device_communicator.all2all_manager
assert all2all_manager is not None
if (prepare_finalize.activation_format ==
FusedMoEActivationFormat.BatchedExperts):
logger.debug("BatchedTritonExperts %s", self.moe)
assert self.moe.dp_size == all2all_manager.dp_world_size
return BatchedTritonExperts(
max_num_tokens=self.moe.max_num_tokens,
world_size=all2all_manager.world_size,
# dp_size actually means tp_size, bug in pplx kernels
dp_size=all2all_manager.tp_group.world_size,
num_dispatchers=prepare_finalize.num_dispatchers(),
)
else:
logger.debug("TritonExperts %s", self.moe)
@@ -652,14 +642,12 @@ class FusedMoE(torch.nn.Module):
get_tensor_model_parallel_world_size())
dp_size_ = (dp_size
if dp_size is not None else get_dp_group().world_size)
world_size_ = get_world_group().world_size
vllm_config = get_current_vllm_config()
self.moe_parallel_config: FusedMoEParallelConfig = (
FusedMoEParallelConfig.make(
tp_size_=tp_size_,
dp_size_=dp_size_,
world_size_=world_size_,
vllm_parallel_config=vllm_config.parallel_config))
self.global_num_experts = num_experts + num_redundant_experts
@@ -1186,9 +1174,9 @@ class FusedMoE(torch.nn.Module):
logical_replica_count: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Route the input hidden states to the top-k experts based on the
Route the input hidden states to the top-k experts based on the
router logits.
Returns:
(topk_weights, topk_ids) (tuple[torch.Tensor, torch.Tensor]):
The weights and *global physical* expert ids of the top-k experts.
@@ -1299,6 +1287,8 @@ class FusedMoE(torch.nn.Module):
topk_ids = topk_ids.to(dtype=indices_type)
assert topk_ids.dtype == indices_type or indices_type is None
return topk_weights, topk_ids
def must_reduce_shared_expert_outputs(self) -> bool:
@@ -193,6 +193,10 @@ class FusedMoEPrepareAndFinalize(ABC):
"""
raise NotImplementedError
@abstractmethod
def num_dispatchers(self) -> int:
raise NotImplementedError
class FusedMoEPermuteExpertsUnpermute(ABC):
"""
@@ -8,7 +8,7 @@ import torch
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
from vllm.model_executor.layers.fused_moe.utils import (
moe_kernel_quantize_input)
_validate_scale_shape, moe_kernel_quantize_input)
from vllm.utils import cdiv, round_up
@@ -32,16 +32,16 @@ def pplx_hidden_dim_scale_bytes(
elem_size = torch.float32.itemsize
if per_act_token_quant:
# per-token
# per-token (M x 1)
assert block_shape is None
hidden_scale_bytes = elem_size
elif block_shape is not None:
# per-group
# per-group (M x K_tiles)
block_size = block_shape[1]
num_blocks = cdiv(hidden_dim, block_size)
hidden_scale_bytes = num_blocks * elem_size
else:
# per-tensor
# per-tensor (1 x 1)
hidden_scale_bytes = elem_size
else:
hidden_dim_bytes = hidden_dim * in_dtype.itemsize
@@ -53,25 +53,22 @@ def pplx_hidden_dim_scale_bytes(
)
# The max_num_tokens, world_size and dp_size must be the same
# as the ones used to create the AllToAll.
class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
def __init__(
self,
a2a: pplx.AllToAll,
max_num_tokens: int,
world_size: int,
rank: int,
dp_size: int,
num_local_experts: int,
num_dispatchers: int,
):
super().__init__()
assert max_num_tokens > 0
assert num_local_experts > 0
self.a2a = a2a
self.max_num_tokens = max_num_tokens
self.world_size = world_size
self.rank = rank
self.dp_size = dp_size
self.num_local_experts = num_local_experts
self.num_dispatchers_ = num_dispatchers
@property
def activation_format(self) -> mk.FusedMoEActivationFormat:
@@ -83,6 +80,9 @@ class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
def topk_indices_dtype(self) -> Optional[torch.dtype]:
return torch.uint32
def num_dispatchers(self) -> int:
return self.num_dispatchers_
def prepare(
self,
a1: torch.Tensor,
@@ -120,42 +120,64 @@ class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
per_act_token_quant=quant_config.per_act_token_quant,
block_shape=quant_config.block_shape)
if a1q_scale is not None:
if a1q_scale.numel() == 1:
orig_a_scale_block_shape = 1
else:
orig_a_scale_block_shape = a1q_scale.shape[-1]
a1q_scale = a1q_scale.repeat(repeat_rows, repeat_cols)
_validate_scale_shape(a1q, a1q_scale, quant_config.per_act_token_quant,
quant_config.block_shape)
# rem_experts need to be 0 for pplx to work properly.
rem_experts = num_experts % self.world_size
assert rem_experts == 0
num_local_experts = ((num_experts // self.world_size) +
(1 if self.rank < rem_experts else 0))
if a1q_scale is not None:
scalar_scales = a1q_scale.numel() == 1
# pplx requires 2-d scales even for scalar scales
if a1q_scale.dim() <= 1:
assert scalar_scales
a1q_scale = a1q_scale.view(1, 1)
orig_a_scale_block_shape = a1q_scale.shape[-1]
if not quant_config.is_block_quantized:
# TODO (bnell): use group_broadcast instead?
a1q_scale = a1q_scale.repeat(repeat_rows, repeat_cols)
assert a1q_scale is None or a1q_scale.ndim == 2, \
f"{0 if a1q_scale is None else (a1q_scale.ndim, a1q_scale.shape)}"
expert_num_tokens = torch.empty(
num_local_experts,
self.num_local_experts,
dtype=torch.int32,
device=device,
)
num_dp = self.world_size // self.dp_size
expert_x = torch.empty(
(num_local_experts, self.max_num_tokens * num_dp, hidden_dim),
(self.num_local_experts,
self.max_num_tokens * self.num_dispatchers(), hidden_dim),
dtype=a1q.dtype,
device=device,
)
expert_x_scale: Optional[torch.Tensor] = None
if a1q.dtype.itemsize == 1:
block_size = (quant_config.block_shape[1]
if quant_config.block_shape is not None else 1)
if quant_config.is_per_act_token:
# (M x 1) -> (E x M x K)
final_dim = expert_x.size(2)
elif quant_config.is_per_tensor:
# (1 x 1) -> (E x 1 x 1)
final_dim = 1
else:
# (M x K_tiles) -> (E x M x K_tiles)
assert quant_config.block_shape is not None
num_blocks = cdiv(expert_x.size(2),
quant_config.block_shape[1])
final_dim = num_blocks
expert_x_scale_shape = (
self.num_local_experts,
expert_x.size(1),
round_up(final_dim, 4) # round up for alignment
)
expert_x_scale = torch.empty(
(num_local_experts, expert_x.size(1),
round_up(
(expert_x.size(2) + block_size - 1) // block_size, 4)),
expert_x_scale_shape,
dtype=torch.float32,
device=device,
device=expert_x.device,
)
# This argument is optional, defaults to indices.size(0)
@@ -171,8 +193,10 @@ class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
indices=topk_ids,
bound_m=bound_m,
)
if expert_x_scale is not None:
expert_x_scale = expert_x_scale[:, :, :orig_a_scale_block_shape]
assert expert_x_scale.ndim == 3
return expert_x, expert_x_scale, expert_num_tokens, None, None
@@ -184,13 +208,16 @@ class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
topk_ids: torch.Tensor,
apply_router_weight_on_input: bool,
) -> None:
num_tokens = output.size(0) # M
# This argument is optional
# There's not much point setting this unless it is != topk_ids.size(0)
bound_m: Optional[torch.Tensor] = None
assert topk_ids.size(0) == num_tokens, (
f"{topk_ids.size(0)} == {num_tokens}")
# TODO (bnell): fails in test_pplx_moe.py, figure out what's going on
#num_tokens = output.size(0) # M
#assert topk_ids.size(0) == num_tokens, (
# f"{topk_ids.size(0)} == {num_tokens}")
assert topk_ids.size() == topk_weights.size(), (
f"{topk_ids.size()} == {topk_weights.size()}")
assert output.size(0) <= self.max_num_tokens, (
f"{output.size(0)} <= {self.max_num_tokens}")
assert output.size(1) == fused_expert_output.size(-1)
@@ -24,6 +24,9 @@ class MoEPrepareAndFinalizeNoEP(mk.FusedMoEPrepareAndFinalize):
def topk_indices_dtype(self) -> Optional[torch.dtype]:
return None
def num_dispatchers(self) -> int:
return 1
def prepare(
self,
a1: torch.Tensor,
+34 -3
View File
@@ -99,9 +99,20 @@ def _fp8_perm(m: torch.Tensor, idx: torch.Tensor) -> torch.Tensor:
return m[idx, ...]
# TODO(bnell): better name
def maybe_fix_scales(scales: Optional[torch.Tensor],
num_experts: int) -> Optional[torch.Tensor]:
def normalize_scales_shape(
scales: Optional[torch.Tensor]) -> Optional[torch.Tensor]:
if scales is not None:
if scales.numel() == 1:
scales = scales.view(1, 1)
else:
scales = scales.view(-1, scales.size(-1))
return scales
def normalize_batched_scales_shape(
scales: Optional[torch.Tensor],
num_experts: int,
) -> Optional[torch.Tensor]:
if scales is not None and scales.ndim < 3:
if scales.numel() == 1:
scales = scales.view(1)
@@ -111,3 +122,23 @@ def maybe_fix_scales(scales: Optional[torch.Tensor],
scales = scales.view(num_experts, -1, scales.size(-1))
return scales
def _validate_scale_shape(
a: torch.Tensor,
a_scale: Optional[torch.Tensor],
per_act_token_quant: bool,
block_shape: Optional[list[int]],
) -> None:
if a_scale is None:
return
if not per_act_token_quant and block_shape is None:
assert a_scale.numel() == 1, f"{a_scale.shape}"
elif per_act_token_quant:
assert a_scale.shape[0] == a.shape[0] and a_scale.shape[1] == 1, (
f"{a_scale.shape[0]} == {a.shape[0]} and {a_scale.shape[1]} == 1")
else:
assert block_shape is not None
expected = (a.shape[0], cdiv(a.shape[1], block_shape[1]))
assert a_scale.shape == expected, f"{a_scale.shape} == {expected}"
@@ -573,6 +573,41 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
from vllm.model_executor.layers.fused_moe import fused_experts
self.fused_experts_func = fused_experts
def select_gemm_impl(
self,
prepare_finalize: FusedMoEPrepareAndFinalize,
moe: FusedMoEConfig,
) -> FusedMoEPermuteExpertsUnpermute:
from vllm.model_executor.layers.fused_moe import TritonExperts
from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
BatchedTritonExperts)
assert not self.rocm_aiter_moe_enabled and not self.use_marlin
logger.debug("BatchedTritonExperts(%s)", self.__class__.__name__)
if (prepare_finalize.activation_format ==
FusedMoEActivationFormat.BatchedExperts):
max_num_tokens_per_rank = prepare_finalize.max_num_tokens_per_rank(
)
assert max_num_tokens_per_rank is not None
return BatchedTritonExperts(
max_num_tokens=max_num_tokens_per_rank,
num_dispatchers=prepare_finalize.num_dispatchers(),
use_fp8_w8a8=True,
block_shape=self.quant_config.weight_block_size,
per_act_token_quant=(
self.input_quant.strategy == QuantizationStrategy.TOKEN),
)
else:
return TritonExperts(
use_fp8_w8a8=True,
block_shape=self.quant_config.weight_block_size,
per_act_token_quant=(
self.input_quant.strategy == QuantizationStrategy.TOKEN),
)
def apply(
self,
layer: torch.nn.Module,
@@ -610,7 +645,9 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
num_expert_group=num_expert_group,
custom_routing_function=custom_routing_function,
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias)
e_score_correction_bias=e_score_correction_bias,
indices_type=self.topk_indices_dtype,
)
if self.rocm_aiter_moe_enabled:
return self.rocm_aiter_fused_experts_func(
@@ -832,18 +869,25 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
use_batched_format = (prepare_finalize.activation_format ==
FusedMoEActivationFormat.BatchedExperts)
num_dispatchers = prepare_finalize.num_dispatchers()
num_experts = (moe.num_local_experts
if use_batched_format else moe.num_experts)
logger.debug("CutlassExpertsFp8(%s)", self.__class__.__name__)
experts = CutlassExpertsFp8(
num_experts,
moe.in_dtype,
self.input_quant.strategy == QuantizationStrategy.TOKEN,
self.weight_quant.strategy == QuantizationStrategy.CHANNEL,
num_dispatchers=num_dispatchers,
use_batched_format=use_batched_format,
)
self.disable_expert_map = not experts.supports_expert_map()
self.disable_expert_map = (num_dispatchers > 1
or not experts.supports_expert_map())
return experts
def apply(
@@ -802,10 +802,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
self.quant_config.weight_block_size, False)
return BatchedTritonOrDeepGemmExperts(
max_num_tokens=max_num_tokens_per_rank,
world_size=prepare_finalize.
world_size, # type: ignore [attr-defined]
dp_size=prepare_finalize.
dp_size, # type: ignore [attr-defined]
num_dispatchers=prepare_finalize.num_dispatchers(),
use_fp8_w8a8=True,
block_shape=self.quant_config.weight_block_size,
per_act_token_quant=False,
+1 -1
View File
@@ -135,7 +135,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
router_logits, _ = self.gate(hidden_states)
final_hidden_states = self.experts(hidden_states=hidden_states,
router_logits=router_logits)
final_hidden_states = final_hidden_states
if self.tp_size > 1:
final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel( # noqa E501
final_hidden_states)