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, 19:30 UTC
[TPU] Reduce compilation time & Upgrade PyTorch XLA version (#6856)
Woosuk Kwon committed
Jul 27, 2024 at 10:28 UTC
fad5576c58864a6c2cf528f67e60e03a949b3dac
6 files changed
+24
-7
Dockerfile.tpu
+1
-1
@@ -1,4 +1,4 @@
1
-ARG NIGHTLY_DATE="20240713"
1
+ARG NIGHTLY_DATE="20240726"
2
ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE"
3
4
FROM $BASE_IMAGE
docs/source/getting_started/tpu-installation.rst
+8
-1
@@ -56,7 +56,7 @@ First, install the dependencies:
56
$ pip uninstall torch torch-xla -y
57
58
$ # Install PyTorch and PyTorch XLA.
59
- $ export DATE="+20240713"
59
+ $ export DATE="+20240726"
60
$ pip install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-nightly${DATE}-cp310-cp310-linux_x86_64.whl
61
$ pip install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-nightly${DATE}-cp310-cp310-linux_x86_64.whl
62
@@ -75,6 +75,13 @@ Next, build vLLM from source. This will only take a few seconds:
75
$ VLLM_TARGET_DEVICE="tpu" python setup.py develop
76
77
78
+.. note::
79
+
80
+ Since TPU relies on XLA which requires static shapes, vLLM bucketizes the possible input shapes and compiles an XLA graph for each different shape.
81
+ The compilation time may take 20~30 minutes in the first run.
82
+ However, the compilation time reduces to ~5 minutes afterwards because the XLA graphs are cached in the disk (in :code:`VLLM_XLA_CACHE_PATH` or :code:`~/.cache/vllm/xla_cache` by default).
83
+
84
+
85
.. tip::
86
87
If you encounter the following error:
vllm/attention/backends/pallas.py
-1
@@ -3,7 +3,6 @@ from typing import Any, Dict, List, Optional, Tuple, Type
3
4
import torch
5
import torch_xla.experimental.custom_kernel # Required to register custom ops.
6
-import torch_xla.experimental.dynamo_set_buffer_donor
6
7
from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
8
AttentionMetadata, AttentionType)
vllm/distributed/device_communicators/tpu_communicator.py
+2
-1
@@ -6,6 +6,7 @@ from vllm.platforms import current_platform
6
7
if current_platform.is_tpu():
8
import torch_xla.core.xla_model as xm
9
+ import torch_xla.runtime as xr
10
from torch_xla._internal import pjrt
11
12
@@ -20,7 +21,7 @@ class TpuCommunicator:
21
local_rank = dist.get_rank(group)
22
world_size = dist.get_world_size(group)
23
pjrt.initialize_multiprocess(local_rank, world_size)
23
- xm._init_world_size_ordinal()
24
+ xr._init_world_size_ordinal()
25
26
def all_reduce(self, x: torch.Tensor) -> torch.Tensor:
27
return xm.all_reduce(xm.REDUCE_SUM, x)
vllm/worker/tpu_model_runner.py
+13
-2
@@ -7,6 +7,7 @@ import numpy as np
7
import torch
8
import torch.nn as nn
9
import torch_xla.core.xla_model as xm
10
+import torch_xla.runtime as xr
11
12
from vllm.attention import AttentionMetadata, get_attn_backend
13
from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, ModelConfig,
@@ -127,7 +128,7 @@ class TPUModelRunner(ModelRunnerBase[ModelInputForTPU]):
128
# determine the order of concatenating the output tensors.
129
# As a workaround, we use the xm's rank assignment only when loading
130
# the embedding weights.
130
- xm_tp_rank = xm.get_ordinal()
131
+ xm_tp_rank = xr.global_ordinal()
132
with patch(
133
"vllm.model_executor.layers.vocab_parallel_embedding."
134
"get_tensor_model_parallel_rank",
@@ -146,7 +147,17 @@ class TPUModelRunner(ModelRunnerBase[ModelInputForTPU]):
147
xm.wait_device_ops()
148
149
model = ModelWrapper(model)
149
- self.model = torch.compile(model, backend="openxla", fullgraph=True)
150
+ # NOTE(woosuk): There are two stages of compilation: torch.compile and
151
+ # XLA compilation. Setting dynamic=True can reduce the torch.compile
152
+ # overhead by reusing the FX graph for different shapes.
153
+ # However, the XLA graph will still require static shapes and needs to
154
+ # be re-compiled for every different shapes. This overhead is inevitable
155
+ # in the first run, but can be skipped afterwards as we cache the XLA
156
+ # graphs in the disk (VLLM_XLA_CACHE_PATH).
157
+ self.model = torch.compile(model,
158
+ backend="openxla",
159
+ fullgraph=True,
160
+ dynamic=True)
161
162
def _dummy_run(
163
self,
vllm/worker/tpu_worker.py
-1
@@ -3,7 +3,6 @@ from typing import List, Optional, Tuple, Union
3
4
import torch
5
import torch_xla.core.xla_model as xm
6
-import torch_xla.experimental.dynamo_set_buffer_donor # noqa: F401
6
import torch_xla.runtime as xr
7
8
import vllm.envs as envs