[TPU] Increase block size and reset block shapes (vllm-project#16458)

Author: bythew3i

examples/offline_inference/tpu.py

@@ -22,7 +22,8 @@ def main():
     # In real workloads, `enforace_eager` should be `False`.
     llm = LLM(model="Qwen/Qwen2-1.5B-Instruct",
               max_num_batched_tokens=64,
-              max_num_seqs=4)
+              max_num_seqs=4,
+              max_model_len=128)
     outputs = llm.generate(prompts, sampling_params)
     print("-" * 50)
     for output, answer in zip(outputs, answers):

requirements/tpu.txt

@@ -18,9 +18,9 @@ setuptools==78.1.0
 --find-links https://storage.googleapis.com/libtpu-releases/index.html
 --find-links https://storage.googleapis.com/jax-releases/jax_nightly_releases.html
 --find-links https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
-torch==2.8.0.dev20250408
-torchvision==0.22.0.dev20250408
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250408-cp39-cp39-linux_x86_64.whl ; python_version == "3.9"
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250408-cp310-cp310-linux_x86_64.whl ; python_version == "3.10"
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250408-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
+torch==2.8.0.dev20250430
+torchvision==0.22.0.dev20250430
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250430-cp39-cp39-linux_x86_64.whl ; python_version == "3.9"
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250430-cp310-cp310-linux_x86_64.whl ; python_version == "3.10"
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250430-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
 

vllm/platforms/tpu.py

@@ -76,9 +76,9 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         from vllm.config import CompilationLevel
 
         cache_config = vllm_config.cache_config
+        # For v0, the default block size is 16.
         if cache_config and cache_config.block_size is None:
             cache_config.block_size = 16
-
         compilation_config = vllm_config.compilation_config
 
         # TPU only supports DYNAMO_ONCE compilation level
@@ -101,16 +101,18 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         if envs.VLLM_USE_V1:
             from vllm.v1.attention.backends.pallas import (
                 PallasAttentionBackend)
+            cache_config.block_size = PallasAttentionBackend.get_page_size(
+                vllm_config)
             min_page_size = PallasAttentionBackend.get_min_page_size(
                 vllm_config)
-            if min_page_size > vllm_config.cache_config.block_size:
+            if min_page_size > cache_config.block_size:
                 logger.warning(
                     "Increase the page size from %s to %s to make sure there's"
                     "no SMEM OOM",
-                    vllm_config.cache_config.block_size,
+                    cache_config.block_size,
                     min_page_size,
                 )
-                vllm_config.cache_config.block_size = min_page_size
+                cache_config.block_size = min_page_size
 
         parallel_config = vllm_config.parallel_config
         scheduler_config = vllm_config.scheduler_config

vllm/utils.py

@@ -707,6 +707,13 @@ def cdiv(a: int, b: int) -> int:
     return -(a // -b)
 
 
+def next_power_of_2(n) -> int:
+    """The next power of 2 (inclusive)"""
+    if n < 1:
+        return 1
+    return 1 << (n - 1).bit_length()
+
+
 def round_up(x: int, y: int) -> int:
     return ((x + y - 1) // y) * y
 

vllm/v1/attention/backends/pallas.py

@@ -12,7 +12,7 @@
 from vllm.attention.backends.utils import CommonAttentionState
 from vllm.config import VllmConfig
 from vllm.logger import init_logger
-from vllm.utils import cdiv
+from vllm.utils import cdiv, next_power_of_2
 
 logger = init_logger(__name__)
 
@@ -65,6 +65,20 @@ def get_min_page_size(vllm_config: VllmConfig) -> int:
         min_page_size = 1 << (min_page_size - 1).bit_length()
         return min_page_size
 
+    # TPU has limited SREGs (scalar registers), if page_size is too small, we
+    # can spill SREGs easily which leads to bad performance. The strategy we
+    # apply here is trying to split max-model-len to 16 pages which make the
+    # spill less likely. Meanwhile we make sure the page size is in [16, 256].
+    @staticmethod
+    def get_page_size(vllm_config: VllmConfig) -> int:
+        page_size = next_power_of_2(
+            vllm_config.model_config.max_model_len) // 16
+        if page_size <= 16:
+            return 16
+        if page_size >= 256:
+            return 256
+        return page_size
+
 
 @dataclass
 class PallasMetadata: