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

[Models] Add remaining model PP support (#7168)

Browse Source 
Signed-off-by: Muralidhar Andoorveedu <muralidhar.andoorveedu@centml.ai>
Signed-off-by: Murali Andoorveedu <muralidhar.andoorveedu@centml.ai>
Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Murali Andoorveedu
2024-10-03 19:56:58 -07:00
committed by GitHub
parent 303d44790a
commit 0f6d7a9a34
69 changed files with 2585 additions and 1344 deletions
Download Patch File Download Diff File Expand all files Collapse all files

63
vllm/model_executor/models/gemma.py
View File

@@ -15,7 +15,7 @@
# limitations under the License.
"""Inference-only Gemma model compatible with HuggingFace weights."""
from functools import lru_cache
from typing import Iterable, List, Optional, Set, Tuple
from typing import Iterable, List, Optional, Set, Tuple, Union
import torch
from torch import nn
@@ -23,7 +23,7 @@ from transformers import GemmaConfig
from vllm.attention import Attention, AttentionMetadata
from vllm.config import CacheConfig, LoRAConfig
from vllm.distributed import get_tensor_model_parallel_world_size
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 GeluAndMul
from vllm.model_executor.layers.layernorm import GemmaRMSNorm
@@ -31,8 +31,7 @@ from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
from vllm.model_executor.layers.vocab_parallel_embedding import (
@@ -41,7 +40,9 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsLoRA
from .interfaces import SupportsLoRA, SupportsPP
from .utils import (is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers)
logger = init_logger(__name__)
@@ -245,6 +246,7 @@ class GemmaModel(nn.Module):
config: GemmaConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
self.config = config
@@ -253,10 +255,11 @@ class GemmaModel(nn.Module):
config.vocab_size,
config.hidden_size,
)
self.layers = nn.ModuleList([
GemmaDecoderLayer(config, cache_config, quant_config)
for _ in range(config.num_hidden_layers)
])
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: GemmaDecoderLayer(config, cache_config, quant_config
),
prefix=f"{prefix}.layers")
self.norm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
# Normalize the embedding by sqrt(hidden_size)
@@ -265,6 +268,9 @@ class GemmaModel(nn.Module):
# See https://github.com/huggingface/transformers/pull/29402
normalizer = self.config.hidden_size**0.5
self.register_buffer("normalizer", torch.tensor(normalizer))
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], config.hidden_size))
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.embed_tokens(input_ids)
@@ -275,29 +281,38 @@ class GemmaModel(nn.Module):
positions: torch.Tensor,
kv_caches: List[torch.Tensor],
attn_metadata: AttentionMetadata,
intermediate_tensors: Optional[IntermediateTensors] = None,
intermediate_tensors: Optional[IntermediateTensors],
inputs_embeds: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if inputs_embeds is not None:
hidden_states = inputs_embeds
) -> Union[torch.Tensor, IntermediateTensors]:
if get_pp_group().is_first_rank:
if inputs_embeds is not None:
hidden_states = inputs_embeds
else:
hidden_states = self.get_input_embeddings(input_ids)
hidden_states *= self.normalizer
residual = None
else:
hidden_states = self.get_input_embeddings(input_ids)
hidden_states *= self.normalizer
residual = None
for i in range(len(self.layers)):
hidden_states = intermediate_tensors["hidden_states"]
residual = intermediate_tensors["residual"]
for i in range(self.start_layer, self.end_layer):
layer = self.layers[i]
hidden_states, residual = layer(
positions,
hidden_states,
kv_caches[i],
kv_caches[i - self.start_layer],
attn_metadata,
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
class GemmaForCausalLM(nn.Module, SupportsLoRA):
class GemmaForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
@@ -339,6 +354,8 @@ class GemmaForCausalLM(nn.Module, SupportsLoRA):
self.model = GemmaModel(config, cache_config, quant_config)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.sampler = Sampler()
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors)
def forward(
self,
@@ -347,9 +364,9 @@ class GemmaForCausalLM(nn.Module, SupportsLoRA):
kv_caches: List[torch.Tensor],
attn_metadata: AttentionMetadata,
intermediate_tensors: Optional[IntermediateTensors] = None,
) -> torch.Tensor:
) -> Union[torch.Tensor, IntermediateTensors]:
hidden_states = self.model(input_ids, positions, kv_caches,
attn_metadata)
attn_metadata, intermediate_tensors)
return hidden_states
def compute_logits(
@@ -388,6 +405,8 @@ class GemmaForCausalLM(nn.Module, SupportsLoRA):
# 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)
@@ -400,6 +419,8 @@ class GemmaForCausalLM(nn.Module, SupportsLoRA):
# 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)