[MOE] Standardize and abstract fused shared expert optimization selection (#51695)
Signed-off-by: Felix Marty <Felix.Marty@amd.com>
This commit is contained in:
@@ -46,6 +46,10 @@ from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoEFactory,
|
||||
fused_moe_make_expert_params_mapping,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.utils import (
|
||||
is_model_fused_shared_expert_compatible,
|
||||
resolve_layer_fused_shared_expert,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
ColumnParallelLinear,
|
||||
@@ -148,10 +152,14 @@ class AXK1MoE(nn.Module):
|
||||
self.n_local_physical_experts = self.n_physical_experts // self.ep_size
|
||||
|
||||
self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
|
||||
self.is_fusion_moe_shared_experts_enabled = (
|
||||
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
)
|
||||
if config.n_shared_experts is None or self.is_fusion_moe_shared_experts_enabled:
|
||||
|
||||
self.is_fused_shared_expert_enabled = False
|
||||
if config.n_shared_experts is not None:
|
||||
self.is_fused_shared_expert_enabled = resolve_layer_fused_shared_expert(
|
||||
quant_config, prefix
|
||||
)
|
||||
|
||||
if config.n_shared_experts is None or self.is_fused_shared_expert_enabled:
|
||||
self.shared_experts = None
|
||||
else:
|
||||
intermediate_size = config.moe_intermediate_size * config.n_shared_experts
|
||||
@@ -189,8 +197,9 @@ class AXK1MoE(nn.Module):
|
||||
num_redundant_experts=self.n_redundant_experts,
|
||||
is_sequence_parallel=self.is_sequence_parallel,
|
||||
n_shared_experts=config.n_shared_experts
|
||||
if self.is_fusion_moe_shared_experts_enabled
|
||||
if self.is_fused_shared_expert_enabled
|
||||
else None,
|
||||
fuse_shared_experts=self.is_fused_shared_expert_enabled,
|
||||
)
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
@@ -725,6 +734,12 @@ class AXK1Model(nn.Module):
|
||||
prefix=f"{prefix}.layers",
|
||||
)
|
||||
|
||||
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
|
||||
self.layers,
|
||||
AXK1MoE,
|
||||
"mlp",
|
||||
)
|
||||
|
||||
if get_pp_group().is_last_rank:
|
||||
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
else:
|
||||
@@ -789,9 +804,6 @@ class AXK1Model(nn.Module):
|
||||
return hidden_states
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
rocm_aiter_moe_shared_expert_enabled = (
|
||||
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
)
|
||||
stacked_params_mapping: list[tuple[str, str, int | str]] = [
|
||||
# (param_name, shard_name, shard_id)
|
||||
("gate_up_proj", "gate_proj", 0),
|
||||
@@ -822,7 +834,7 @@ class AXK1Model(nn.Module):
|
||||
num_experts=self.config.n_routed_experts
|
||||
+ (
|
||||
(self.config.n_shared_experts or 0)
|
||||
if rocm_aiter_moe_shared_expert_enabled
|
||||
if self.is_fused_shared_expert_enabled
|
||||
else 0
|
||||
),
|
||||
num_redundant_experts=self.num_redundant_experts,
|
||||
@@ -839,7 +851,7 @@ class AXK1Model(nn.Module):
|
||||
continue # skip spec decode layers for main model
|
||||
|
||||
is_fusion_moe_shared_experts_layer = (
|
||||
rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
self.is_fused_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
)
|
||||
|
||||
for param_name, weight_name, shard_id in stacked_params_mapping:
|
||||
|
||||
@@ -7,13 +7,15 @@ import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.distributed import tensor_model_parallel_all_gather
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
fused_moe_make_expert_params_mapping,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.utils import (
|
||||
is_model_fused_shared_expert_compatible,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
@@ -255,6 +257,11 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
|
||||
self.moe_mlp_layers.append(layer.mlp)
|
||||
self.moe_layers.append(layer.mlp.experts)
|
||||
self.extract_moe_parameters(example_moe)
|
||||
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
|
||||
self.model.layers.values(),
|
||||
DeepseekV2MoE,
|
||||
"mtp_block.mlp",
|
||||
)
|
||||
|
||||
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.model.embed_input_ids(input_ids)
|
||||
@@ -285,9 +292,6 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
|
||||
return self.model.compute_logits(hidden_states, spec_step_idx)
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
rocm_aiter_moe_shared_expert_enabled = (
|
||||
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
)
|
||||
stacked_params_mapping = [
|
||||
("gate_up_proj", "gate_proj", 0),
|
||||
("gate_up_proj", "up_proj", 1),
|
||||
@@ -310,7 +314,7 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
|
||||
num_experts=self.config.n_routed_experts
|
||||
+ (
|
||||
self.config.n_shared_experts
|
||||
if rocm_aiter_moe_shared_expert_enabled
|
||||
if self.is_fused_shared_expert_enabled
|
||||
else 0
|
||||
),
|
||||
)
|
||||
@@ -326,7 +330,7 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
|
||||
if spec_layer is None:
|
||||
continue
|
||||
is_fusion_moe_shared_experts_layer = (
|
||||
rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
self.is_fused_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
)
|
||||
name = self._rewrite_spec_layer_name(spec_layer, name)
|
||||
|
||||
|
||||
@@ -54,6 +54,10 @@ from vllm.model_executor.layers.fused_moe import (
|
||||
GateLinear,
|
||||
fused_moe_make_expert_params_mapping,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.utils import (
|
||||
is_model_fused_shared_expert_compatible,
|
||||
resolve_layer_fused_shared_expert,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
ColumnParallelLinear,
|
||||
@@ -329,9 +333,13 @@ class DeepseekV2MoE(nn.Module):
|
||||
self.n_local_physical_experts = self.n_physical_experts // self.ep_size
|
||||
|
||||
self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
|
||||
self.is_fusion_moe_shared_experts_enabled = (
|
||||
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
)
|
||||
|
||||
self.is_fused_shared_expert_enabled = False
|
||||
if config.n_shared_experts is not None:
|
||||
self.is_fused_shared_expert_enabled = resolve_layer_fused_shared_expert(
|
||||
quant_config, prefix
|
||||
)
|
||||
|
||||
if (
|
||||
self.is_rocm_aiter_moe_enabled
|
||||
and self.gate.e_score_correction_bias is not None
|
||||
@@ -340,7 +348,7 @@ class DeepseekV2MoE(nn.Module):
|
||||
# Accumulates in fp32; avoids bf16->fp32 cast.
|
||||
self.gate.set_out_dtype(self.gate.weight.dtype)
|
||||
|
||||
if config.n_shared_experts is None or self.is_fusion_moe_shared_experts_enabled:
|
||||
if config.n_shared_experts is None or self.is_fused_shared_expert_enabled:
|
||||
self.shared_experts = None
|
||||
else:
|
||||
intermediate_size = config.moe_intermediate_size * config.n_shared_experts
|
||||
@@ -377,8 +385,9 @@ class DeepseekV2MoE(nn.Module):
|
||||
is_sequence_parallel=self.is_sequence_parallel,
|
||||
reduce_results=reduce_results,
|
||||
n_shared_experts=config.n_shared_experts
|
||||
if self.is_fusion_moe_shared_experts_enabled
|
||||
if self.is_fused_shared_expert_enabled
|
||||
else None,
|
||||
fuse_shared_experts=self.is_fused_shared_expert_enabled,
|
||||
router_logits_dtype=self.gate.out_dtype,
|
||||
)
|
||||
|
||||
@@ -1401,6 +1410,12 @@ class DeepseekV2Model(nn.Module):
|
||||
prefix=f"{prefix}.layers",
|
||||
)
|
||||
|
||||
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
|
||||
self.layers,
|
||||
DeepseekV2MoE,
|
||||
"mlp",
|
||||
)
|
||||
|
||||
if get_pp_group().is_last_rank:
|
||||
self.norm = RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
|
||||
else:
|
||||
@@ -1511,9 +1526,6 @@ class DeepseekV2Model(nn.Module):
|
||||
return hidden_states
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
rocm_aiter_moe_shared_expert_enabled = (
|
||||
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
)
|
||||
stacked_params_mapping: list[tuple[str, str, int | str]] = [
|
||||
# (param_name, shard_name, shard_id)
|
||||
("gate_up_proj", "gate_proj", 0),
|
||||
@@ -1554,7 +1566,7 @@ class DeepseekV2Model(nn.Module):
|
||||
num_experts=self.config.n_routed_experts
|
||||
+ (
|
||||
self.config.n_shared_experts
|
||||
if rocm_aiter_moe_shared_expert_enabled
|
||||
if self.is_fused_shared_expert_enabled
|
||||
else 0
|
||||
),
|
||||
num_redundant_experts=self.num_redundant_experts,
|
||||
@@ -1582,7 +1594,7 @@ class DeepseekV2Model(nn.Module):
|
||||
continue # this layer has no indexer; drop its checkpoint weights
|
||||
|
||||
is_fusion_moe_shared_experts_layer = (
|
||||
rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
self.is_fused_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
)
|
||||
|
||||
if _try_load_fp8_indexer_wk(
|
||||
|
||||
@@ -39,12 +39,15 @@ from vllm.distributed import (
|
||||
get_pp_group,
|
||||
get_tensor_model_parallel_world_size,
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoEFactory,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.utils import (
|
||||
is_model_fused_shared_expert_compatible,
|
||||
resolve_layer_fused_shared_expert,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
MergedColumnParallelLinear,
|
||||
@@ -72,8 +75,6 @@ from .utils import (
|
||||
skip_spec_layers,
|
||||
)
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class Glm4MoeMLP(nn.Module):
|
||||
def __init__(
|
||||
@@ -159,14 +160,15 @@ class Glm4MoE(nn.Module):
|
||||
self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
|
||||
self.n_local_physical_experts = self.n_physical_experts // self.ep_size
|
||||
|
||||
# AITER fused shared-expert (FSE) gate; mirrors the deepseek_v2.py
|
||||
# pattern (see Glm4MoE / MoERunner wiring there).
|
||||
self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
|
||||
self.is_fusion_moe_shared_experts_enabled = (
|
||||
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
)
|
||||
|
||||
if config.n_shared_experts is None or self.is_fusion_moe_shared_experts_enabled:
|
||||
self.is_fused_shared_expert_enabled = False
|
||||
if config.n_shared_experts is not None:
|
||||
self.is_fused_shared_expert_enabled = resolve_layer_fused_shared_expert(
|
||||
quant_config, prefix
|
||||
)
|
||||
|
||||
if config.n_shared_experts is None or self.is_fused_shared_expert_enabled:
|
||||
self.shared_experts = None
|
||||
else:
|
||||
intermediate_size = config.moe_intermediate_size * config.n_shared_experts
|
||||
@@ -200,10 +202,9 @@ class Glm4MoE(nn.Module):
|
||||
num_redundant_experts=self.n_redundant_experts,
|
||||
router_logits_dtype=torch.float32,
|
||||
n_shared_experts=(
|
||||
config.n_shared_experts
|
||||
if self.is_fusion_moe_shared_experts_enabled
|
||||
else None
|
||||
config.n_shared_experts if self.is_fused_shared_expert_enabled else None
|
||||
),
|
||||
fuse_shared_experts=self.is_fused_shared_expert_enabled,
|
||||
)
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
@@ -443,6 +444,12 @@ class Glm4MoeModel(nn.Module):
|
||||
prefix=f"{prefix}.layers",
|
||||
)
|
||||
|
||||
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
|
||||
self.layers,
|
||||
Glm4MoE,
|
||||
"mlp",
|
||||
)
|
||||
|
||||
if get_pp_group().is_last_rank:
|
||||
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
else:
|
||||
@@ -488,6 +495,7 @@ class Glm4MoeModel(nn.Module):
|
||||
skip_spec_layers(weights, self.config),
|
||||
n_routed_experts=self.config.n_routed_experts,
|
||||
n_shared_experts=self.config.n_shared_experts or 1,
|
||||
enabled=self.is_fused_shared_expert_enabled,
|
||||
)
|
||||
loader = AutoWeightsLoader(self)
|
||||
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
|
||||
|
||||
@@ -34,7 +34,6 @@ from torch import nn
|
||||
if TYPE_CHECKING:
|
||||
from transformers.models.glm4_moe_lite import Glm4MoeLiteConfig
|
||||
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.distributed import (
|
||||
@@ -44,6 +43,9 @@ from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
fused_moe_make_expert_params_mapping,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.utils import (
|
||||
is_model_fused_shared_expert_compatible,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
@@ -254,6 +256,12 @@ class Glm4MoeLiteModel(nn.Module):
|
||||
prefix=f"{prefix}.layers",
|
||||
)
|
||||
|
||||
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
|
||||
self.layers,
|
||||
Glm4MoeLite,
|
||||
"mlp",
|
||||
)
|
||||
|
||||
if get_pp_group().is_last_rank:
|
||||
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
else:
|
||||
@@ -320,9 +328,6 @@ class Glm4MoeLiteModel(nn.Module):
|
||||
)
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
rocm_aiter_moe_shared_expert_enabled = (
|
||||
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
)
|
||||
stacked_params_mapping = [
|
||||
# (param_name, shard_name, shard_id)
|
||||
("gate_up_proj", "gate_proj", 0),
|
||||
@@ -345,7 +350,7 @@ class Glm4MoeLiteModel(nn.Module):
|
||||
num_experts=self.config.n_routed_experts
|
||||
+ (
|
||||
self.config.n_shared_experts
|
||||
if rocm_aiter_moe_shared_expert_enabled
|
||||
if self.is_fused_shared_expert_enabled
|
||||
else 0
|
||||
),
|
||||
)
|
||||
@@ -361,7 +366,7 @@ class Glm4MoeLiteModel(nn.Module):
|
||||
continue # skip spec decode layers for main model
|
||||
|
||||
is_fusion_moe_shared_experts_layer = (
|
||||
rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
self.is_fused_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
)
|
||||
|
||||
for param_name, weight_name, shard_id in stacked_params_mapping:
|
||||
|
||||
@@ -30,12 +30,14 @@ import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
MoERunner,
|
||||
fused_moe_make_expert_params_mapping,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.utils import (
|
||||
is_model_fused_shared_expert_compatible,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
@@ -224,6 +226,11 @@ class Glm4MoeLiteMTP(nn.Module, SupportsPP, Glm4MixtureOfExperts):
|
||||
self.moe_mlp_layers.append(layer.mlp)
|
||||
self.moe_layers.append(layer.mlp.experts)
|
||||
self.extract_moe_parameters(example_moe)
|
||||
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
|
||||
self.model.layers.values(),
|
||||
Glm4MoeLite,
|
||||
"mtp_block.mlp",
|
||||
)
|
||||
|
||||
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.model.embed_input_ids(input_ids)
|
||||
@@ -250,9 +257,6 @@ class Glm4MoeLiteMTP(nn.Module, SupportsPP, Glm4MixtureOfExperts):
|
||||
return self.model.compute_logits(hidden_states, spec_step_idx)
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
rocm_aiter_moe_shared_expert_enabled = (
|
||||
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
)
|
||||
stacked_params_mapping = [
|
||||
("gate_up_proj", "gate_proj", 0),
|
||||
("gate_up_proj", "up_proj", 1),
|
||||
@@ -268,7 +272,7 @@ class Glm4MoeLiteMTP(nn.Module, SupportsPP, Glm4MixtureOfExperts):
|
||||
num_experts=self.config.n_routed_experts
|
||||
+ (
|
||||
self.config.n_shared_experts
|
||||
if rocm_aiter_moe_shared_expert_enabled
|
||||
if self.is_fused_shared_expert_enabled
|
||||
else 0
|
||||
),
|
||||
)
|
||||
@@ -282,7 +286,7 @@ class Glm4MoeLiteMTP(nn.Module, SupportsPP, Glm4MixtureOfExperts):
|
||||
if spec_layer is None:
|
||||
continue
|
||||
is_fusion_moe_shared_experts_layer = (
|
||||
rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
self.is_fused_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
)
|
||||
name = self._rewrite_spec_layer_name(spec_layer, name)
|
||||
for param_name, weight_name, shard_id in stacked_params_mapping:
|
||||
|
||||
@@ -31,12 +31,14 @@ import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
from vllm.config import CacheConfig, ParallelConfig, VllmConfig
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
MoERunner,
|
||||
fused_moe_make_expert_params_mapping,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.utils import (
|
||||
is_model_fused_shared_expert_compatible,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
@@ -212,6 +214,11 @@ class Glm4MoeMTP(nn.Module, Glm4MixtureOfExperts):
|
||||
self.moe_mlp_layers.append(layer.mlp)
|
||||
self.moe_layers.append(layer.mlp.experts)
|
||||
self.extract_moe_parameters(example_moe)
|
||||
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
|
||||
self.model.layers.values(),
|
||||
Glm4MoE,
|
||||
"mtp_block.mlp",
|
||||
)
|
||||
|
||||
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.model.embed_input_ids(input_ids)
|
||||
@@ -238,10 +245,6 @@ class Glm4MoeMTP(nn.Module, Glm4MixtureOfExperts):
|
||||
return self.model.compute_logits(hidden_states, spec_step_idx)
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
# FSE weight loading mirrors glm4_moe.py / deepseek_mtp.py.
|
||||
rocm_aiter_moe_shared_expert_enabled = (
|
||||
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
)
|
||||
stacked_params_mapping = [
|
||||
# (param_name, shard_name, shard_id)
|
||||
("qkv_proj", "q_proj", "q"),
|
||||
@@ -254,7 +257,7 @@ class Glm4MoeMTP(nn.Module, Glm4MixtureOfExperts):
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
num_experts = self.config.n_routed_experts
|
||||
if rocm_aiter_moe_shared_expert_enabled and self.config.n_shared_experts:
|
||||
if self.is_fused_shared_expert_enabled and self.config.n_shared_experts:
|
||||
num_experts += self.config.n_shared_experts
|
||||
expert_params_mapping = fused_moe_make_expert_params_mapping(
|
||||
self,
|
||||
@@ -279,7 +282,7 @@ class Glm4MoeMTP(nn.Module, Glm4MixtureOfExperts):
|
||||
name = self._rewrite_spec_layer_name(spec_layer, name)
|
||||
|
||||
is_fusion_moe_shared_experts_layer = (
|
||||
rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
self.is_fused_shared_expert_enabled and ("mlp.shared_experts" in name)
|
||||
)
|
||||
|
||||
for param_name, weight_name, shard_id in stacked_params_mapping:
|
||||
|
||||
@@ -29,13 +29,15 @@ from collections.abc import Iterable
|
||||
import torch
|
||||
from torch import nn
|
||||
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.distributed import (
|
||||
get_pp_group,
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe.utils import (
|
||||
is_model_fused_shared_expert_compatible,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import GemmaRMSNorm as Qwen3_5RMSNorm
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.mamba.gdn.qwen_gdn_linear_attn import (
|
||||
@@ -78,7 +80,6 @@ from .qwen3_next import (
|
||||
Qwen3NextModel,
|
||||
Qwen3NextSparseMoeBlock,
|
||||
QwenNextMixtureOfExperts,
|
||||
_is_shared_expert_fse_compatible,
|
||||
)
|
||||
from .qwen3_vl import (
|
||||
Qwen3_VisionTransformer,
|
||||
@@ -256,6 +257,11 @@ class Qwen3_5Model(Qwen3NextModel):
|
||||
self.start_layer, self.end_layer, self.layers = make_layers(
|
||||
config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers"
|
||||
)
|
||||
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
|
||||
self.layers,
|
||||
Qwen3NextSparseMoeBlock,
|
||||
"mlp",
|
||||
)
|
||||
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], config.hidden_size
|
||||
)
|
||||
@@ -274,8 +280,7 @@ class Qwen3_5Model(Qwen3NextModel):
|
||||
if "moe" in self.config.model_type:
|
||||
weights = maybe_fuse_shared_experts(
|
||||
weights,
|
||||
enabled=rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
and _is_shared_expert_fse_compatible(self.quant_config),
|
||||
enabled=self.is_fused_shared_expert_enabled,
|
||||
n_routed_experts=self.config.num_experts,
|
||||
n_shared_experts=1,
|
||||
ckpt_prefix="mlp.shared_expert",
|
||||
|
||||
@@ -7,11 +7,13 @@ from collections.abc import Iterable
|
||||
import torch
|
||||
from torch import nn
|
||||
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import VllmConfig, get_current_vllm_config
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.distributed import get_pp_group, tensor_model_parallel_all_gather
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe.utils import (
|
||||
is_model_fused_shared_expert_compatible,
|
||||
)
|
||||
from vllm.model_executor.layers.linear import ColumnParallelLinear
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
@@ -25,8 +27,8 @@ from vllm.model_executor.models.qwen3_5 import (
|
||||
Qwen3_5RMSNorm,
|
||||
)
|
||||
from vllm.model_executor.models.qwen3_next import (
|
||||
Qwen3NextSparseMoeBlock,
|
||||
QwenNextMixtureOfExperts,
|
||||
_is_shared_expert_fse_compatible,
|
||||
)
|
||||
from vllm.model_executor.models.utils import sequence_parallel_chunk
|
||||
from vllm.sequence import IntermediateTensors
|
||||
@@ -122,6 +124,11 @@ class Qwen3_5MultiTokenPredictor(nn.Module):
|
||||
for idx in range(self.num_mtp_layers)
|
||||
)
|
||||
vllm_config.quant_config = original_quant
|
||||
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
|
||||
self.layers,
|
||||
Qwen3NextSparseMoeBlock,
|
||||
"mlp",
|
||||
)
|
||||
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], config.hidden_size
|
||||
)
|
||||
@@ -185,10 +192,7 @@ class Qwen3_5MultiTokenPredictor(nn.Module):
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
weights = maybe_fuse_shared_experts(
|
||||
weights,
|
||||
enabled=rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
and _is_shared_expert_fse_compatible(
|
||||
get_current_vllm_config().quant_config
|
||||
),
|
||||
enabled=self.is_fused_shared_expert_enabled,
|
||||
n_routed_experts=getattr(self.config, "num_experts", 0),
|
||||
n_shared_experts=1,
|
||||
ckpt_prefix="mlp.shared_expert",
|
||||
|
||||
@@ -8,7 +8,6 @@ from itertools import islice
|
||||
import torch
|
||||
from torch import nn
|
||||
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, ModelConfig, VllmConfig
|
||||
from vllm.distributed import (
|
||||
@@ -18,9 +17,12 @@ from vllm.distributed import (
|
||||
tensor_model_parallel_all_gather,
|
||||
tensor_model_parallel_reduce_scatter,
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoEFactory
|
||||
from vllm.model_executor.layers.fused_moe.utils import (
|
||||
is_model_fused_shared_expert_compatible,
|
||||
resolve_layer_fused_shared_expert,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_qk_norm_rope import fused_qk_rmsnorm_rope_gate
|
||||
from vllm.model_executor.layers.layernorm import (
|
||||
GemmaRMSNorm as Qwen3NextRMSNorm,
|
||||
@@ -73,8 +75,6 @@ from .utils import (
|
||||
maybe_prefix,
|
||||
)
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
KVCache = tuple[torch.Tensor, torch.Tensor]
|
||||
|
||||
|
||||
@@ -90,26 +90,6 @@ def _should_use_sequence_parallel(vllm_config: VllmConfig) -> bool:
|
||||
)
|
||||
|
||||
|
||||
def _is_shared_expert_fse_compatible(quant_config) -> bool:
|
||||
"""Check if shared expert can be fused with routed experts.
|
||||
|
||||
FSE requires that shared and routed expert weights use the same
|
||||
quantization format. Returns False when the shared expert is
|
||||
excluded from quantization (e.g. float32 shared in an MXFP4 model)
|
||||
or has a different quant spec than routed experts.
|
||||
"""
|
||||
if quant_config is None:
|
||||
return True
|
||||
# Quark stores its full config dict in quant_config.quant_config
|
||||
raw_config = getattr(quant_config, "quant_config", None)
|
||||
if not isinstance(raw_config, dict):
|
||||
return True
|
||||
exclude = raw_config.get("exclude", [])
|
||||
if not exclude:
|
||||
return True
|
||||
return not any("shared_expert." in str(e) for e in exclude)
|
||||
|
||||
|
||||
class Qwen3NextSparseMoeBlock(nn.Module):
|
||||
def __init__(self, vllm_config: VllmConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
@@ -157,15 +137,18 @@ class Qwen3NextSparseMoeBlock(nn.Module):
|
||||
prefix=f"{prefix}.shared_expert_gate",
|
||||
)
|
||||
|
||||
_fse_requested = rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
|
||||
_fse_enabled = _fse_requested and _is_shared_expert_fse_compatible(quant_config)
|
||||
if _fse_requested and not _fse_enabled:
|
||||
logger.warning(
|
||||
"VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS is enabled but "
|
||||
"shared expert has a different quantization spec than routed "
|
||||
"experts. Falling back to non-fused shared expert path."
|
||||
self.is_fused_shared_expert_enabled = False
|
||||
if config.shared_expert_intermediate_size > 0:
|
||||
self.is_fused_shared_expert_enabled = resolve_layer_fused_shared_expert(
|
||||
quant_config,
|
||||
prefix,
|
||||
shared_expert_name="shared_expert",
|
||||
)
|
||||
if _fse_enabled or config.shared_expert_intermediate_size <= 0:
|
||||
|
||||
if (
|
||||
self.is_fused_shared_expert_enabled
|
||||
or config.shared_expert_intermediate_size <= 0
|
||||
):
|
||||
self.shared_expert = None
|
||||
else:
|
||||
self.shared_expert = Qwen3NextMLP(
|
||||
@@ -193,6 +176,7 @@ class Qwen3NextSparseMoeBlock(nn.Module):
|
||||
num_redundant_experts=self.n_redundant_experts,
|
||||
is_sequence_parallel=self.is_sequence_parallel,
|
||||
n_shared_experts=1 if self.shared_expert is None else None,
|
||||
fuse_shared_experts=self.is_fused_shared_expert_enabled,
|
||||
shared_expert_gate=self.shared_expert_gate
|
||||
if self.shared_expert is None
|
||||
else None,
|
||||
@@ -599,6 +583,11 @@ class Qwen3NextModel(nn.Module, EagleModelMixin):
|
||||
self.start_layer, self.end_layer, self.layers = make_layers(
|
||||
config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers"
|
||||
)
|
||||
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
|
||||
self.layers,
|
||||
Qwen3NextSparseMoeBlock,
|
||||
"mlp",
|
||||
)
|
||||
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], config.hidden_size
|
||||
)
|
||||
@@ -678,6 +667,7 @@ class Qwen3NextModel(nn.Module, EagleModelMixin):
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
weights = maybe_fuse_shared_experts(
|
||||
weights,
|
||||
enabled=self.is_fused_shared_expert_enabled,
|
||||
n_routed_experts=getattr(self.config, "num_experts", 0),
|
||||
n_shared_experts=1,
|
||||
ckpt_prefix="mlp.shared_expert",
|
||||
|
||||
@@ -11,6 +11,9 @@ from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.distributed import get_pp_group, tensor_model_parallel_all_gather
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe.utils import (
|
||||
is_model_fused_shared_expert_compatible,
|
||||
)
|
||||
from vllm.model_executor.layers.linear import ColumnParallelLinear
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
@@ -21,6 +24,7 @@ from vllm.model_executor.models.qwen3_next import (
|
||||
Qwen3NextDecoderLayer,
|
||||
Qwen3NextModel,
|
||||
Qwen3NextRMSNorm,
|
||||
Qwen3NextSparseMoeBlock,
|
||||
QwenNextMixtureOfExperts,
|
||||
)
|
||||
from vllm.model_executor.models.utils import sequence_parallel_chunk
|
||||
@@ -91,6 +95,11 @@ class Qwen3NextMultiTokenPredictor(nn.Module):
|
||||
for idx in range(self.num_mtp_layers)
|
||||
)
|
||||
|
||||
self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible(
|
||||
self.layers,
|
||||
Qwen3NextSparseMoeBlock,
|
||||
"mlp",
|
||||
)
|
||||
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], config.hidden_size
|
||||
)
|
||||
@@ -155,6 +164,7 @@ class Qwen3NextMultiTokenPredictor(nn.Module):
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
weights = maybe_fuse_shared_experts(
|
||||
weights,
|
||||
enabled=self.is_fused_shared_expert_enabled,
|
||||
n_routed_experts=self.config.num_experts,
|
||||
n_shared_experts=1,
|
||||
ckpt_prefix="mlp.shared_expert",
|
||||
|
||||
Reference in New Issue
Block a user