[GLM-OCR] GLM-OCR with MTP Support (#33005)

Signed-off-by: zRzRzRzRzRzRzR <2448370773@qq.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Yuxuan Zhang
2026-01-26 06:24:43 -08:00
committed by GitHub
co-authored by Isotr0py
parent dcd80206b7
commit bb17e8f11c
14 changed files with 873 additions and 8 deletions
+99 -2
View File
@@ -39,13 +39,22 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader,
maybe_remap_kv_scale_name,
)
from vllm.sequence import IntermediateTensors
from vllm.v1.attention.backend import AttentionType
from .interfaces import SupportsLoRA, SupportsPP
from .llama import LlamaMLP as Glm4MLP
from .llama import LlamaModel
from .utils import AutoWeightsLoader, PPMissingLayer, maybe_prefix
from .utils import (
AutoWeightsLoader,
PPMissingLayer,
is_pp_missing_parameter,
maybe_prefix,
)
class Glm4Attention(nn.Module):
@@ -78,7 +87,15 @@ class Glm4Attention(nn.Module):
# Number of KV heads is less than TP size, so we replicate
# the KV heads across multiple tensor parallel GPUs.
assert tp_size % self.total_num_kv_heads == 0
config.rope_parameters.setdefault("partial_rotary_factor", 0.5)
rope_params = getattr(config, "rope_parameters", None)
if isinstance(rope_params, dict) and "partial_rotary_factor" in rope_params:
config.rope_parameters.setdefault(
"partial_rotary_factor", rope_params["partial_rotary_factor"]
)
else:
config.rope_parameters.setdefault("partial_rotary_factor", 0.5)
self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
self.head_dim = head_dim or hidden_size // self.total_num_heads
self.q_size = self.num_heads * self.head_dim
@@ -220,6 +237,73 @@ class Glm4Model(LlamaModel):
vllm_config=vllm_config, prefix=prefix, layer_type=Glm4DecoderLayer
)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
stacked_params_mapping = [
# (param_name, shard_name, shard_id)
(".qkv_proj", ".q_proj", "q"),
(".qkv_proj", ".k_proj", "k"),
(".qkv_proj", ".v_proj", "v"),
(".gate_up_proj", ".gate_proj", 0),
(".gate_up_proj", ".up_proj", 1),
]
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
for name, loaded_weight in weights:
spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
if spec_layer is not None:
continue
if "rotary_emb.inv_freq" in name:
continue
if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
# Models trained using ColossalAI may include these tensors in
# the checkpoint. Skip them.
continue
if self.quant_config is not None and (
scale_name := self.quant_config.get_cache_scale(name)
):
# Loading kv cache quantization scales
param = params_dict[scale_name]
weight_loader = getattr(param, "weight_loader", default_weight_loader)
loaded_weight = (
loaded_weight if loaded_weight.dim() == 0 else loaded_weight[0]
)
weight_loader(param, loaded_weight)
loaded_params.add(scale_name)
continue
if "scale" in name or "zero_point" in name:
# Remapping the name of FP8 kv-scale or zero point.
name = maybe_remap_kv_scale_name(name, params_dict)
if name is None:
continue
for param_name, weight_name, shard_id in stacked_params_mapping:
if weight_name not in name:
continue
name = name.replace(weight_name, param_name)
# Skip loading extra bias for GPTQ models.
if name.endswith(".bias") and name not in params_dict:
continue
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
weight_loader = param.weight_loader
weight_loader(param, loaded_weight, shard_id)
break
else:
# Skip loading extra bias for GPTQ models.
if name.endswith(".bias") and name not in params_dict:
continue
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
weight_loader = getattr(param, "weight_loader", default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params
class Glm4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
packed_modules_mapping = {
@@ -293,3 +377,16 @@ class Glm4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
)
return loader.load_weights(weights)
def get_spec_layer_idx_from_weight_name(
config: Glm4Config, weight_name: str
) -> int | None:
if hasattr(config, "num_nextn_predict_layers") and (
config.num_nextn_predict_layers > 0
):
layer_idx = config.num_hidden_layers
for i in range(config.num_nextn_predict_layers):
if f"layers.{layer_idx + i}." in weight_name:
return layer_idx + i
return None