A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 09:30 UTC
[torch.compile] Adding torch compile annotations to some models (#9639)
Signed-off-by: youkaichao <youkaichao@gmail.com> Co-authored-by: youkaichao <youkaichao@gmail.com>
Yongzao committed
Oct 24, 2024 at 15:54 UTC
8a02cd045ac661481ba2672846e09f5b57110f40
7 files changed
+13
-3
docs/source/models/supported_models.rst
+1
-1
@@ -144,7 +144,7 @@ Text Generation
144
- ✅︎
145
* - :code:`JAISLMHeadModel`
146
- Jais
147
- - :code:`core42/jais-13b`, :code:`core42/jais-13b-chat`, :code:`core42/jais-30b-v3`, :code:`core42/jais-30b-chat-v3`, etc.
147
+ - :code:`inceptionai/jais-13b`, :code:`inceptionai/jais-13b-chat`, :code:`inceptionai/jais-30b-v3`, :code:`inceptionai/jais-30b-chat-v3`, etc.
148
-
149
- ✅︎
150
* - :code:`JambaForCausalLM`
tests/distributed/test_pipeline_parallel.py
+1
-1
@@ -145,7 +145,7 @@ TEXT_GENERATION_MODELS = {
145
# Uses Llama
146
# "internlm/internlm-chat-7b": PPTestSettings.fast(),
147
"internlm/internlm2-chat-7b": PPTestSettings.fast(trust_remote_code=True),
148
- "core42/jais-13b-chat": PPTestSettings.fast(),
148
+ "inceptionai/jais-13b-chat": PPTestSettings.fast(),
149
# TODO: Implement PP
150
# "ai21labs/AI21-Jamba-1.5-Mini": PPTestSettings.fast(),
151
"meta-llama/Meta-Llama-3-8B": PPTestSettings.detailed(),
vllm/model_executor/models/jais.py
+3
-1
@@ -1,6 +1,6 @@
1
# coding=utf-8
2
# Adapted from
3
-# https://huggingface.co/core42/jais-30b-chat-v3/blob/main/modeling_jais.py
3
+# https://huggingface.co/inceptionai/jais-30b-chat-v3/blob/main/modeling_jais.py
4
# Copyright 2023 The vLLM team.
5
# Copyright 2023 the Jais authors and HuggingFace Inc. team. All rights
6
# reserved.
@@ -26,6 +26,7 @@ import torch
26
from torch import nn
27
28
from vllm.attention import Attention, AttentionMetadata
29
+from vllm.compilation.decorators import support_torch_compile
30
from vllm.config import CacheConfig
31
from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
32
get_tensor_model_parallel_world_size)
@@ -212,6 +213,7 @@ class JAISBlock(nn.Module):
213
return hidden_states
214
215
216
+@support_torch_compile
217
class JAISModel(nn.Module):
218
219
def __init__(
vllm/model_executor/models/minicpm.py
+2
@@ -29,6 +29,7 @@ from torch import nn
29
from transformers import PretrainedConfig
30
31
from vllm.attention import Attention, AttentionMetadata
32
+from vllm.compilation.decorators import support_torch_compile
33
from vllm.config import CacheConfig, LoRAConfig
34
from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
35
get_tensor_model_parallel_world_size,
@@ -348,6 +349,7 @@ class MiniCPMDecoderLayer(nn.Module):
349
return hidden_states, None
350
351
352
+@support_torch_compile
353
class MiniCPMModel(nn.Module):
354
355
def __init__(
vllm/model_executor/models/mpt.py
+2
@@ -7,6 +7,7 @@ import torch
7
import torch.nn as nn
8
9
from vllm.attention import Attention, AttentionMetadata
10
+from vllm.compilation.decorators import support_torch_compile
11
from vllm.config import CacheConfig
12
from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
13
get_tensor_model_parallel_world_size)
@@ -204,6 +205,7 @@ class MPTBlock(nn.Module):
205
return hidden_states
206
207
208
+@support_torch_compile
209
class MPTModel(nn.Module):
210
211
def __init__(
vllm/model_executor/models/nemotron.py
+2
@@ -27,6 +27,7 @@ import torch
27
from torch import nn
28
29
from vllm.attention import Attention, AttentionMetadata
30
+from vllm.compilation.decorators import support_torch_compile
31
from vllm.config import CacheConfig, LoRAConfig
32
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
33
from vllm.model_executor.layers.activation import get_act_fn
@@ -290,6 +291,7 @@ class NemotronDecoderLayer(nn.Module):
291
return hidden_states, residual
292
293
294
+@support_torch_compile
295
class NemotronModel(nn.Module):
296
297
def __init__(
vllm/model_executor/models/olmo.py
+2
@@ -28,6 +28,7 @@ from torch import nn
28
from transformers import OlmoConfig
29
30
from vllm.attention import Attention, AttentionMetadata
31
+from vllm.compilation.decorators import support_torch_compile
32
from vllm.config import CacheConfig
33
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
34
from vllm.model_executor.layers.activation import SiluAndMul
@@ -221,6 +222,7 @@ class OlmoDecoderLayer(nn.Module):
222
return hidden_states
223
224
225
+@support_torch_compile
226
class OlmoModel(nn.Module):
227
228
def __init__(self,