[TPU] Reduce compilation time & Upgrade PyTorch XLA version (#6856)

Author: WoosukKwon

Dockerfile.tpu

@@ -1,4 +1,4 @@
-ARG NIGHTLY_DATE="20240713"
+ARG NIGHTLY_DATE="20240726"
 ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE"
 
 FROM $BASE_IMAGE

docs/source/getting_started/tpu-installation.rst

@@ -56,7 +56,7 @@ First, install the dependencies:
     $ pip uninstall torch torch-xla -y
 
     $ # Install PyTorch and PyTorch XLA.
-    $ export DATE="+20240713"
+    $ export DATE="+20240726"
     $ pip install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-nightly${DATE}-cp310-cp310-linux_x86_64.whl
     $ pip install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-nightly${DATE}-cp310-cp310-linux_x86_64.whl
 
@@ -75,6 +75,13 @@ Next, build vLLM from source. This will only take a few seconds:
     $ VLLM_TARGET_DEVICE="tpu" python setup.py develop
 
 
+.. note::
+
+    Since TPU relies on XLA which requires static shapes, vLLM bucketizes the possible input shapes and compiles an XLA graph for each different shape.
+    The compilation time may take 20~30 minutes in the first run.
+    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).
+
+
 .. tip::
 
     If you encounter the following error:

vllm/attention/backends/pallas.py

@@ -3,7 +3,6 @@
 
 import torch
 import torch_xla.experimental.custom_kernel  # Required to register custom ops.
-import torch_xla.experimental.dynamo_set_buffer_donor
 
 from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               AttentionMetadata, AttentionType)

vllm/distributed/device_communicators/tpu_communicator.py

@@ -6,6 +6,7 @@
 
 if current_platform.is_tpu():
     import torch_xla.core.xla_model as xm
+    import torch_xla.runtime as xr
     from torch_xla._internal import pjrt
 
 
@@ -20,7 +21,7 @@ def __init__(self, group: ProcessGroup):
         local_rank = dist.get_rank(group)
         world_size = dist.get_world_size(group)
         pjrt.initialize_multiprocess(local_rank, world_size)
-        xm._init_world_size_ordinal()
+        xr._init_world_size_ordinal()
 
     def all_reduce(self, x: torch.Tensor) -> torch.Tensor:
         return xm.all_reduce(xm.REDUCE_SUM, x)

vllm/worker/tpu_model_runner.py

@@ -7,6 +7,7 @@
 import torch
 import torch.nn as nn
 import torch_xla.core.xla_model as xm
+import torch_xla.runtime as xr
 
 from vllm.attention import AttentionMetadata, get_attn_backend
 from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, ModelConfig,
@@ -127,7 +128,7 @@ def load_model(self) -> None:
         # determine the order of concatenating the output tensors.
         # As a workaround, we use the xm's rank assignment only when loading
         # the embedding weights.
-        xm_tp_rank = xm.get_ordinal()
+        xm_tp_rank = xr.global_ordinal()
         with patch(
                 "vllm.model_executor.layers.vocab_parallel_embedding."
                 "get_tensor_model_parallel_rank",
@@ -146,7 +147,17 @@ def load_model(self) -> None:
         xm.wait_device_ops()
 
         model = ModelWrapper(model)
-        self.model = torch.compile(model, backend="openxla", fullgraph=True)
+        # NOTE(woosuk): There are two stages of compilation: torch.compile and
+        # XLA compilation. Setting dynamic=True can reduce the torch.compile
+        # overhead by reusing the FX graph for different shapes.
+        # However, the XLA graph will still require static shapes and needs to
+        # be re-compiled for every different shapes. This overhead is inevitable
+        # in the first run, but can be skipped afterwards as we cache the XLA
+        # graphs in the disk (VLLM_XLA_CACHE_PATH).
+        self.model = torch.compile(model,
+                                   backend="openxla",
+                                   fullgraph=True,
+                                   dynamic=True)
 
     def _dummy_run(
         self,

vllm/worker/tpu_worker.py

@@ -3,7 +3,6 @@
 
 import torch
 import torch_xla.core.xla_model as xm
-import torch_xla.experimental.dynamo_set_buffer_donor  # noqa: F401
 import torch_xla.runtime as xr
 
 import vllm.envs as envs