biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model] use AutoWeightsLoader for granite, granitemoe, granitemoeshared, grok1, mixtral (#16325)

Browse Source 
Signed-off-by: Aaron Ang <aaron.angyd@gmail.com>
This commit is contained in:
Aaron Ang
2025-04-09 23:07:40 -04:00
committed by GitHub
parent 1da6a09274
commit a564797151
5 changed files with 366 additions and 336 deletions
Download Patch File Download Diff File Expand all files Collapse all files

143
vllm/model_executor/models/granite.py
View File

@@ -50,8 +50,8 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsLoRA, SupportsPP
from .utils import (PPMissingLayer, is_pp_missing_parameter, make_layers,
maybe_prefix)
from .utils import (AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter,
make_layers, maybe_prefix)
class GraniteMLP(nn.Module):
@@ -260,6 +260,7 @@ class GraniteModel(nn.Module):
lora_config = vllm_config.lora_config
self.config = config
self.quant_config = quant_config
lora_vocab = (lora_config.lora_extra_vocab_size *
(lora_config.max_loras or 1)) if lora_config else 0
self.vocab_size = config.vocab_size + lora_vocab
@@ -321,6 +322,65 @@ class GraniteModel(nn.Module):
hidden_states = self.norm(hidden_states)
return hidden_states
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:
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
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
# Remapping the name of FP8 kv-scale.
name = maybe_remap_kv_scale_name(name, params_dict)
if name is None:
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 GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
packed_modules_mapping = {
@@ -428,71 +488,18 @@ class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
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),
skip_prefixes = [
"rotary_emb.inv_freq",
# Models trained using ColossalAI may include these tensors in
# the checkpoint. Skip them.
"rotary_emb.cos_cached",
"rotary_emb.sin_cached",
]
params_dict = dict(self.named_parameters())
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
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
# With tie_word_embeddings, we can skip lm_head.weight
# The weight might appear unnecessarily in the files if the model is
# processed with quantization, LoRA, fine-tuning, etc.
if self.config.tie_word_embeddings and "lm_head.weight" in name:
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
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
# With tie_word_embeddings, we can skip lm_head.weight
# The weight might appear unnecessarily in the files if the model is
# processed with quantization, LoRA, fine-tuning, etc.
if self.config.tie_word_embeddings:
skip_prefixes.append("lm_head.weight")
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
# Remapping the name of FP8 kv-scale.
name = maybe_remap_kv_scale_name(name, params_dict)
if name is None:
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
loader = AutoWeightsLoader(self, skip_prefixes=skip_prefixes)
return loader.load_weights(weights)