[Model] Use upstream GLM-5.3 and Qwen4-Exp configs and processor (#57387)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-10-02 02:25:18 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent 866f02db8a
commit e3cae8d2ac
36 changed files with 366 additions and 2385 deletions
+1 -3
View File
@@ -8,6 +8,7 @@ from typing import cast
import torch
from torch import nn
from transformers import Qwen4ExpTextConfig
from vllm.config import VllmConfig
from vllm.forward_context import get_forward_context
@@ -15,9 +16,6 @@ from vllm.model_executor.layers.layernorm import GemmaRMSNorm
from vllm.model_executor.layers.linear import ReplicatedLinear
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding.mrope import triton_mrope
from vllm.transformers_utils.configs.qwen4_exp import (
Qwen4ExpTextConfig,
)
from ..common.qsa_cache import (
QSACompressedKeyCache,
+10 -37
View File
@@ -7,6 +7,7 @@ from itertools import islice
import torch
from torch import nn
from transformers import Qwen4ExpConfig, Qwen4ExpTextConfig
from vllm.compilation.decorators import support_torch_compile
from vllm.config import VllmConfig
@@ -47,7 +48,6 @@ from vllm.model_executor.models.qwen3_5 import (
)
from vllm.model_executor.models.qwen3_next import (
Qwen3NextAttention,
Qwen3NextMLP,
Qwen3NextSparseMoeBlock,
)
from vllm.model_executor.models.qwen3_vl import (
@@ -71,13 +71,9 @@ from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import MultiModalFeatureSpec
from vllm.sequence import IntermediateTensors
from vllm.tokenizers.registry import cached_tokenizer_from_config
from vllm.transformers_utils.configs.qwen4_exp import (
Qwen4ExpTextConfig,
)
from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
from vllm.v1.kv_cache_interface import MambaSpec
from ..config import ATTENTION_LAYER_TYPES, QSA_LAYER_TYPE, Qwen4ExpConfig
from .hyperconnection import GatedResidual, HyperConnectionConfig
from .low_latency_gemm import enable_qwen4_exp_low_latency_gemm
from .ple_layer import Qwen4ExpPLELayer
@@ -179,8 +175,6 @@ class Qwen4ExpDecoderLayer(nn.Module):
) -> None:
super().__init__()
config: Qwen4ExpTextConfig = vllm_config.model_config.hf_text_config
model_config = vllm_config.model_config
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
self.config = config
@@ -213,16 +207,13 @@ class Qwen4ExpDecoderLayer(nn.Module):
prefix=f"{prefix}.linear_attn",
gqa_interleaved_layout=False,
)
elif layer_type in ATTENTION_LAYER_TYPES:
use_qsa = (
layer_type == QSA_LAYER_TYPE
or getattr(config, "indexer_n_heads", None) is not None
)
elif layer_type == "qwen_sparse_attention":
use_qsa = getattr(config, "indexer_n_heads", None) is not None
if not use_qsa:
self.self_attn = Qwen3NextAttention(
config,
model_config=model_config,
cache_config=cache_config,
model_config=vllm_config.model_config,
cache_config=vllm_config.cache_config,
quant_config=quant_config,
prefix=f"{prefix}.self_attn",
)
@@ -237,24 +228,9 @@ class Qwen4ExpDecoderLayer(nn.Module):
else:
raise ValueError(f"Invalid layer_type {layer_type}")
mlp_only_layers = getattr(config, "mlp_only_layers", [])
num_experts = getattr(config, "num_experts", 0) or 0
absolute_layer_id = self.layer_idx + 1
is_moe_layer = self.layer_idx not in mlp_only_layers and (
num_experts > 0 and absolute_layer_id % config.decoder_sparse_step == 0
self.mlp = Qwen4ExpSparseMoeBlock(
vllm_config=vllm_config, prefix=f"{prefix}.mlp"
)
if is_moe_layer:
self.mlp = Qwen4ExpSparseMoeBlock(
vllm_config=vllm_config, prefix=f"{prefix}.mlp"
)
else:
self.mlp = Qwen3NextMLP(
hidden_size=config.hidden_size,
intermediate_size=config.intermediate_size,
hidden_act=config.hidden_act,
quant_config=quant_config,
prefix=f"{prefix}.mlp",
)
hc_config = HyperConnectionConfig(
hc_count=config.hc_count,
@@ -313,7 +289,7 @@ class Qwen4ExpDecoderLayer(nn.Module):
if self.layer_type == "linear_attention":
attn_out = self.linear_attn(hidden_states=block_input)
elif self.layer_type in ATTENTION_LAYER_TYPES:
elif self.layer_type == "qwen_sparse_attention":
attn_out = self.self_attn(
hidden_states=block_input,
positions=positions,
@@ -404,11 +380,8 @@ class Qwen4ExpModel(nn.Module):
self._qsa_layer_ids = frozenset(
layer_idx
for layer_idx, layer_type in enumerate(config.layer_types)
if layer_type == QSA_LAYER_TYPE
or (
layer_type == "full_attention"
and getattr(config, "indexer_n_heads", None) is not None
)
if layer_type == "qwen_sparse_attention"
and getattr(config, "indexer_n_heads", None) is not None
)
self.embed_tokens = VocabParallelEmbedding(self.vocab_size, config.hidden_size)
+2 -4
View File
@@ -18,6 +18,7 @@ from collections.abc import Iterable
import regex as re
import torch
from torch import nn
from transformers import Qwen4ExpTextConfig
from vllm.compilation.decorators import support_torch_compile
from vllm.config import VllmConfig, replace, set_current_vllm_config
@@ -45,9 +46,6 @@ from vllm.model_executor.models.utils import (
maybe_prefix,
)
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs.qwen4_exp import (
Qwen4ExpTextConfig,
)
from .hyperconnection import GatedResidual, HyperConnectionConfig
from .low_latency_gemm import enable_qwen4_exp_low_latency_gemm
@@ -222,7 +220,7 @@ class Qwen4ExpMultiTokenPredictor(nn.Module):
self.layers = nn.ModuleList(
Qwen4ExpDecoderLayer(
draft_vllm_config,
layer_type="full_attention",
layer_type="qwen_sparse_attention",
prefix=f"{prefix}.layers.{self.mtp_start_layer_idx + idx}",
)
for idx in range(self.num_mtp_layers)
+1 -3
View File
@@ -8,6 +8,7 @@ from collections.abc import Iterable, Sequence
import torch
import torch.nn.functional as F
from torch import nn
from transformers import Qwen4ExpTextConfig
from vllm.config import CacheConfig, ModelConfig, VllmConfig, get_current_vllm_config
from vllm.forward_context import get_forward_context
@@ -23,9 +24,6 @@ from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig,
)
from vllm.model_executor.models.utils import AutoWeightsLoader
from vllm.transformers_utils.configs.qwen4_exp import (
Qwen4ExpTextConfig,
)
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
from vllm.v1.attention.backends.short_conv_attn import (
+1 -3
View File
@@ -8,6 +8,7 @@ from typing import ClassVar, cast
import torch
from torch import nn
from transformers import Qwen4ExpTextConfig
from vllm.config import VllmConfig
from vllm.config.cache import CacheDType
@@ -23,9 +24,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.models.qwen3_next import Qwen3NextAttention
from vllm.platforms import current_platform
from vllm.transformers_utils.configs.qwen4_exp import (
Qwen4ExpTextConfig,
)
from vllm.utils.torch_utils import (
LayerNameType,
_encode_layer_name,