wehub-resource-sync
302 lines
10 KiB
Python
302 lines
10 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
"""Inference-only Qwen3_5 MTP model."""
|
|
|
|
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.distributed.parallel_state import get_pp_group
|
|
from vllm.logger import init_logger
|
|
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 (
|
|
ParallelLMHead,
|
|
VocabParallelEmbedding,
|
|
)
|
|
from vllm.model_executor.models.interfaces import LocalArgmaxMixin
|
|
from vllm.model_executor.models.qwen3_5 import (
|
|
Qwen3_5DecoderLayer,
|
|
Qwen3_5Model,
|
|
Qwen3_5RMSNorm,
|
|
)
|
|
from vllm.model_executor.models.qwen3_next import (
|
|
QwenNextMixtureOfExperts,
|
|
_is_shared_expert_fse_compatible,
|
|
)
|
|
from vllm.sequence import IntermediateTensors
|
|
from vllm.transformers_utils.configs.qwen3_5 import Qwen3_5TextConfig
|
|
from vllm.transformers_utils.configs.qwen3_5_moe import Qwen3_5MoeTextConfig
|
|
|
|
from .interfaces import (
|
|
MultiModalEmbeddings,
|
|
SupportsMultiModal,
|
|
_require_is_multimodal,
|
|
)
|
|
from .utils import (
|
|
AutoWeightsLoader,
|
|
PPMissingLayer,
|
|
WeightsMapper,
|
|
_merge_multimodal_embeddings,
|
|
make_empty_intermediate_tensors_factory,
|
|
maybe_prefix,
|
|
)
|
|
|
|
logger = init_logger(__name__)
|
|
|
|
|
|
@support_torch_compile(
|
|
dynamic_arg_dims={
|
|
"input_ids": 0,
|
|
# positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl,
|
|
# otherwise (seq_len, ).
|
|
"positions": -1,
|
|
"intermediate_tensors": 0,
|
|
"inputs_embeds": 0,
|
|
"hidden_states": 0,
|
|
}
|
|
)
|
|
class Qwen3_5MultiTokenPredictor(nn.Module):
|
|
hf_to_vllm_mapper = Qwen3_5Model.hf_to_vllm_mapper
|
|
|
|
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
|
super().__init__()
|
|
|
|
model_config = vllm_config.model_config
|
|
quant_config = vllm_config.quant_config
|
|
|
|
config: Qwen3_5TextConfig | Qwen3_5MoeTextConfig = model_config.hf_text_config
|
|
|
|
self.config = config
|
|
|
|
self.vocab_size = config.vocab_size
|
|
|
|
self.mtp_start_layer_idx = config.num_hidden_layers
|
|
self.num_mtp_layers = getattr(config, "mtp_num_hidden_layers", 1)
|
|
|
|
self.embed_tokens = VocabParallelEmbedding(
|
|
self.vocab_size,
|
|
config.hidden_size,
|
|
)
|
|
|
|
# Workaround: mtp.fc is stored as BF16 in NVFP4 checkpoints but is
|
|
# missing from hf_quant_config.json exclude_modules. Force unquantized.
|
|
# Ref: https://github.com/vllm-project/vllm/pull/38650
|
|
# Ref: https://github.com/NVIDIA/Model-Optimizer/pull/1124
|
|
fc_quant = (
|
|
None
|
|
if (quant_config and quant_config.get_name() == "modelopt_fp4")
|
|
else quant_config
|
|
)
|
|
self.fc = ColumnParallelLinear(
|
|
self.config.hidden_size * 2,
|
|
self.config.hidden_size,
|
|
gather_output=True,
|
|
bias=False,
|
|
return_bias=False,
|
|
quant_config=fc_quant,
|
|
prefix=f"{prefix}.fc",
|
|
)
|
|
|
|
self.layers = torch.nn.ModuleList(
|
|
Qwen3_5DecoderLayer(
|
|
vllm_config,
|
|
layer_type="full_attention",
|
|
prefix=f"{prefix}.layers.{idx}",
|
|
)
|
|
for idx in range(self.num_mtp_layers)
|
|
)
|
|
|
|
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
|
["hidden_states", "residual"], config.hidden_size
|
|
)
|
|
|
|
self.norm = Qwen3_5RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
self.pre_fc_norm_hidden = Qwen3_5RMSNorm(
|
|
config.hidden_size, eps=config.rms_norm_eps
|
|
)
|
|
self.pre_fc_norm_embedding = Qwen3_5RMSNorm(
|
|
config.hidden_size, eps=config.rms_norm_eps
|
|
)
|
|
|
|
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
|
return self.embed_tokens(input_ids)
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
intermediate_tensors: IntermediateTensors | None = None,
|
|
inputs_embeds: torch.Tensor | None = None,
|
|
spec_step_idx: int = 0,
|
|
) -> torch.Tensor:
|
|
if get_pp_group().is_first_rank:
|
|
if inputs_embeds is None:
|
|
inputs_embeds = self.embed_input_ids(input_ids)
|
|
assert hidden_states.shape[-1] == inputs_embeds.shape[-1]
|
|
inputs_embeds = self.pre_fc_norm_embedding(inputs_embeds)
|
|
hidden_states = self.pre_fc_norm_hidden(hidden_states)
|
|
hidden_states = torch.cat([inputs_embeds, hidden_states], dim=-1)
|
|
hidden_states = self.fc(hidden_states)
|
|
residual = None
|
|
else:
|
|
assert intermediate_tensors is not None
|
|
hidden_states = intermediate_tensors["hidden_states"]
|
|
residual = intermediate_tensors["residual"]
|
|
|
|
current_step_idx = spec_step_idx % self.num_mtp_layers
|
|
hidden_states, residual = self.layers[current_step_idx](
|
|
positions=positions,
|
|
hidden_states=hidden_states,
|
|
residual=residual,
|
|
)
|
|
|
|
if not get_pp_group().is_last_rank:
|
|
return IntermediateTensors(
|
|
{"hidden_states": hidden_states, "residual": residual}
|
|
)
|
|
|
|
hidden_states, _ = self.norm(hidden_states, residual)
|
|
return hidden_states
|
|
|
|
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
|
mapper = self.hf_to_vllm_mapper
|
|
is_fse = rocm_aiter_ops.is_fusion_moe_shared_experts_enabled() and (
|
|
_is_shared_expert_fse_compatible(get_current_vllm_config().quant_config)
|
|
)
|
|
if is_fse:
|
|
num_routed = getattr(self.config, "num_experts", 0)
|
|
mapper = mapper | WeightsMapper(
|
|
orig_to_new_substr={"mlp.shared_expert.": f"mlp.experts.{num_routed}."}
|
|
)
|
|
loader = AutoWeightsLoader(self)
|
|
return loader.load_weights(weights, mapper=mapper)
|
|
|
|
|
|
@support_torch_compile(
|
|
dynamic_arg_dims={
|
|
"input_ids": 0,
|
|
# positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl,
|
|
# otherwise (seq_len, ).
|
|
"positions": -1,
|
|
"intermediate_tensors": 0,
|
|
"inputs_embeds": 0,
|
|
"hidden_states": 0,
|
|
}
|
|
)
|
|
class Qwen3_5MTP(LocalArgmaxMixin, nn.Module, SupportsMultiModal):
|
|
packed_modules_mapping = {
|
|
"qkv_proj": [
|
|
"q_proj",
|
|
"k_proj",
|
|
"v_proj",
|
|
],
|
|
"gate_up_proj": ["gate_proj", "up_proj"],
|
|
}
|
|
|
|
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
|
config = vllm_config.model_config.hf_text_config
|
|
self.vllm_config = vllm_config
|
|
cache_config = vllm_config.cache_config
|
|
if cache_config.mamba_cache_mode == "all":
|
|
raise NotImplementedError(
|
|
"Qwen3_5MTP currently does not support 'all' prefix caching, "
|
|
"please use '--mamba-cache-mode=align' instead"
|
|
)
|
|
|
|
self.quant_config = vllm_config.quant_config
|
|
|
|
super().__init__()
|
|
self.config = config
|
|
self.model = Qwen3_5MultiTokenPredictor(
|
|
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "mtp")
|
|
)
|
|
|
|
if get_pp_group().is_last_rank:
|
|
if config.tie_word_embeddings:
|
|
self.lm_head = self.model.embed_tokens
|
|
else:
|
|
self.lm_head = ParallelLMHead(
|
|
config.vocab_size,
|
|
config.hidden_size,
|
|
quant_config=self.quant_config,
|
|
prefix=maybe_prefix(prefix, "lm_head"),
|
|
)
|
|
else:
|
|
self.lm_head = PPMissingLayer()
|
|
|
|
self.logits_processor = LogitsProcessor(config.vocab_size)
|
|
|
|
def embed_input_ids(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
multimodal_embeddings: MultiModalEmbeddings | None = None,
|
|
*,
|
|
is_multimodal: torch.Tensor | None = None,
|
|
) -> torch.Tensor:
|
|
inputs_embeds = self._embed_text_input_ids(
|
|
input_ids,
|
|
self.model.embed_input_ids,
|
|
is_multimodal=is_multimodal,
|
|
)
|
|
|
|
if multimodal_embeddings is None or len(multimodal_embeddings) == 0:
|
|
return inputs_embeds
|
|
|
|
is_multimodal = _require_is_multimodal(is_multimodal)
|
|
|
|
inputs_embeds = _merge_multimodal_embeddings(
|
|
inputs_embeds=inputs_embeds,
|
|
multimodal_embeddings=multimodal_embeddings,
|
|
is_multimodal=is_multimodal,
|
|
)
|
|
|
|
return inputs_embeds
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
intermediate_tensors: IntermediateTensors | None = None,
|
|
inputs_embeds: torch.Tensor | None = None,
|
|
**kwargs: object,
|
|
):
|
|
hidden_states = self.model(
|
|
input_ids, positions, hidden_states, intermediate_tensors, inputs_embeds
|
|
)
|
|
return hidden_states
|
|
|
|
def compute_logits(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
spec_step_idx: int = 0,
|
|
) -> torch.Tensor | None:
|
|
return self.logits_processor(self.lm_head, hidden_states)
|
|
|
|
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
|
def remap_weight_names(weights):
|
|
for name, weight in weights:
|
|
if name.startswith("mtp."):
|
|
name = name.replace("mtp.", "model.")
|
|
elif any(key in name for key in ["embed_tokens", "lm_head"]):
|
|
if "embed_tokens" in name:
|
|
name = name.replace("language_model.", "")
|
|
else:
|
|
continue
|
|
yield name, weight
|
|
|
|
loader = AutoWeightsLoader(self)
|
|
return loader.load_weights(remap_weight_names(weights))
|
|
|
|
|
|
class Qwen3_5MoeMTP(Qwen3_5MTP, QwenNextMixtureOfExperts):
|
|
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
|
super().__init__(vllm_config=vllm_config, prefix=prefix)
|
|
self.set_moe_parameters()
|