[TPU] Optimize kv cache update kernel (#20415)

Signed-off-by: Yifei Teng <tengyifei88@gmail.com>

vllm/utils/__init__.py

@@ -947,6 +947,13 @@ def next_power_of_2(n) -> int:
     return 1 << (n - 1).bit_length()
 
 
+def prev_power_of_2(n: int) -> int:
+    """The previous power of 2 (inclusive)"""
+    if n <= 0:
+        return 0
+    return 1 << (n.bit_length() - 1)
+
+
 def round_up(x: int, y: int) -> int:
     return ((x + y - 1) // y) * y
 

vllm/v1/attention/backends/pallas.py

@@ -324,3 +324,9 @@ def kv_cache_update_op_non_xla(kv: torch.Tensor, slot_mapping: torch.Tensor,
                                page_size: int,
                                num_slices_per_block: int) -> torch.Tensor:
     return kv_cache
+
+
+def get_page_size_bytes(block_size: int, num_kv_heads: int, head_size: int,
+                        kv_cache_dtype: torch.dtype) -> int:
+    """Returns the size in bytes of one page of the KV cache."""
+    return block_size * num_kv_heads * head_size * kv_cache_dtype.itemsize

vllm/v1/worker/tpu_model_runner.py

@@ -31,9 +31,10 @@ from vllm.multimodal.inputs import (BatchedTensorInputs, MultiModalKwargs,
 from vllm.multimodal.utils import group_mm_inputs_by_modality
 from vllm.sequence import IntermediateTensors
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, LayerBlockType, cdiv,
-                        is_pin_memory_available)
+                        is_pin_memory_available, prev_power_of_2)
 from vllm.v1.attention.backends.pallas import (PallasAttentionBackend,
-                                               PallasMetadata)
+                                               PallasMetadata,
+                                               get_page_size_bytes)
 from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
 from vllm.v1.kv_cache_interface import (AttentionSpec, FullAttentionSpec,
                                         KVCacheConfig, KVCacheSpec,
@@ -56,8 +57,6 @@ logger = init_logger(__name__)
 INVALID_TOKEN_ID = -1
 # Smallest output size
 MIN_NUM_SEQS = 8
-# Block size used for kv cache updating kernel
-NUM_SLICES_PER_KV_CACHE_UPDATE_BLOCK = 8
 
 
 #########################################################
@@ -139,7 +138,11 @@ class TPUModelRunner(LoRAModelRunnerMixin):
         self.pin_memory = is_pin_memory_available()
         self.dtype = self.model_config.dtype
         if cache_config.cache_dtype == "auto":
-            self.kv_cache_dtype = self.dtype
+            model_dtype = self.dtype
+            if isinstance(model_dtype, str):
+                self.kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[model_dtype]
+            else:
+                self.kv_cache_dtype = model_dtype
         else:
             self.kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
                 cache_config.cache_dtype]
@@ -192,6 +195,14 @@ class TPUModelRunner(LoRAModelRunnerMixin):
         self.max_num_encoder_input_tokens = encoder_compute_budget
         self.encoder_cache_size = encoder_cache_size
 
+        self._num_slices_per_kv_cache_update_block = \
+            _get_num_slices_per_kv_cache_update_block(get_page_size_bytes(
+                block_size=self.block_size,
+                num_kv_heads=self.num_kv_heads,
+                head_size=self.head_size,
+                kv_cache_dtype=self.kv_cache_dtype,
+            ))
+
         # Lazy initialization
         self.model: nn.Module  # Set after load_model
         self.kv_caches: list[torch.Tensor] = []
@@ -719,7 +730,7 @@ class TPUModelRunner(LoRAModelRunnerMixin):
         num_kv_update_slices = slot_mapping_metadata.shape[0]
         padded_num_slices = _get_padded_num_kv_cache_update_slices(
             padded_total_num_scheduled_tokens, self.max_num_reqs,
-            self.block_size)
+            self.block_size, self._num_slices_per_kv_cache_update_block)
         slot_mapping_metadata = np.pad(
             slot_mapping_metadata,
             [[0, padded_num_slices - len(slot_mapping_metadata)], [0, 0]],
@@ -750,8 +761,8 @@ class TPUModelRunner(LoRAModelRunnerMixin):
             num_kv_update_slices=torch.tensor([num_kv_update_slices],
                                               dtype=torch.int32,
                                               device=self.device),
-            num_slices_per_kv_cache_update_block=
-            NUM_SLICES_PER_KV_CACHE_UPDATE_BLOCK,
+            num_slices_per_kv_cache_update_block=self.
+            _num_slices_per_kv_cache_update_block,
         )
         # NOTE(woosuk): Due to chunked prefills, there can be at most 1 partial
         # request in the batch. While we should not sample any token from this
@@ -1197,7 +1208,8 @@ class TPUModelRunner(LoRAModelRunnerMixin):
         position_ids = torch.zeros(num_tokens,
                                    dtype=torch.int32).to(self.device)
         padded_num_slices = _get_padded_num_kv_cache_update_slices(
-            num_tokens, self.max_num_reqs, self.block_size)
+            num_tokens, self.max_num_reqs, self.block_size,
+            self._num_slices_per_kv_cache_update_block)
         num_kv_update_slices = torch.tensor([padded_num_slices],
                                             dtype=torch.int32).to(self.device)
         slot_mapping = torch.zeros((3, padded_num_slices),
@@ -1220,8 +1232,8 @@ class TPUModelRunner(LoRAModelRunnerMixin):
             query_start_loc=query_start_loc,
             num_seqs=num_seqs,
             num_kv_update_slices=num_kv_update_slices,
-            num_slices_per_kv_cache_update_block=
-            NUM_SLICES_PER_KV_CACHE_UPDATE_BLOCK,
+            num_slices_per_kv_cache_update_block=self.
+            _num_slices_per_kv_cache_update_block,
         )
 
         if self.is_multimodal_model:
@@ -1826,19 +1838,41 @@ def _get_padded_token_len(paddings: list[int], x: int) -> int:
     return paddings[index]
 
 
-def _get_padded_num_kv_cache_update_slices(num_tokens: int, max_num_reqs: int,
-                                           page_size: int) -> int:
+def _get_padded_num_kv_cache_update_slices(
+        num_tokens: int, max_num_reqs: int, page_size: int,
+        num_slices_per_kv_cache_update_block: int) -> int:
     """Calculates the padded number of KV cache update slices to avoid
     recompilation."""
     padded_num_slices = 2 * max_num_reqs + num_tokens // page_size
     padded_num_slices = min(padded_num_slices, num_tokens)
     padded_num_slices = (
-        padded_num_slices + NUM_SLICES_PER_KV_CACHE_UPDATE_BLOCK - 1
-    ) // NUM_SLICES_PER_KV_CACHE_UPDATE_BLOCK * \
-        NUM_SLICES_PER_KV_CACHE_UPDATE_BLOCK
+        padded_num_slices + num_slices_per_kv_cache_update_block - 1
+    ) // num_slices_per_kv_cache_update_block * \
+        num_slices_per_kv_cache_update_block
     return padded_num_slices
 
 
+def _get_num_slices_per_kv_cache_update_block(page_size_bytes: int) -> int:
+    """Find the optimum number of slices to copy per Pallas program instance.
+
+    Increasing the number of slices copied in one instance of the kernel program
+    will increase HBM bandwidth utilization via more in-flight DMAs.
+
+    However, it will also use more VMEM, and experimentally, we observed
+    performance regression at 128 slices on v6e, likely due to running
+    out of scalar registers. Thus this function will limit the number of
+    slices to 64.
+    """
+    # Conservative VMEM usage limit: 32 MiB
+    vmem_limit = 32 * 1024 * 1024
+    num_slices_per_block = vmem_limit // page_size_bytes
+    assert num_slices_per_block > 0, "Number of slices should be positive"
+    num_slices_per_block = prev_power_of_2(num_slices_per_block)
+    if num_slices_per_block > 64:
+        num_slices_per_block = 64
+    return num_slices_per_block
+
+
 def replace_set_lora(model):
 
     def _tpu_set_lora(