tpu-inference 0.11.1.dev202511150811__py3-none-any.whl → 0.11.1.dev202512030818__py3-none-any.whl

tpu_inference/runner/compilation_manager.py

@@ -1,13 +1,13 @@
-import os
 import time
-from typing import TYPE_CHECKING, Any, Callable, List, Optional, Tuple
+from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple
 
 import jax
 import jax.numpy as jnp
 import numpy as np
-import vllm.envs as envs
+import vllm.envs as vllm_envs
 from jax.sharding import NamedSharding, PartitionSpec
 
+import tpu_inference.envs as envs
 from tpu_inference.core.disagg_utils import is_disagg_enabled
 from tpu_inference.layers.common.attention_metadata import AttentionMetadata
 from tpu_inference.layers.common.sharding import ShardingAxisName
@@ -15,6 +15,8 @@ from tpu_inference.layers.jax.sample.sampling import sample
 from tpu_inference.layers.jax.sample.sampling_metadata import \
     TPUSupportedSamplingMetadata
 from tpu_inference.logger import init_logger
+from tpu_inference.models.jax.jax_intermediate_tensor import \
+    JaxIntermediateTensors
 from tpu_inference.utils import device_array
 
 if TYPE_CHECKING:
@@ -30,10 +32,10 @@ class CompilationManager:
 
     def __init__(self, runner: "TPUModelRunner"):
         self.runner = runner
-        if not envs.VLLM_DISABLE_COMPILE_CACHE:
+        if not vllm_envs.VLLM_DISABLE_COMPILE_CACHE:
             logger.info("Enabling JAX compile cache.")
             jax.config.update("jax_compilation_cache_dir",
-                              envs.VLLM_XLA_CACHE_PATH)
+                              vllm_envs.VLLM_XLA_CACHE_PATH)
 
     def _create_dummy_tensor(self,
                              shape: Tuple[int, ...],
@@ -67,8 +69,7 @@ class CompilationManager:
         logger.info("Compilation finished in %.2f [secs].", end - start)
 
     def capture_model(self) -> None:
-        if os.getenv("SKIP_JAX_PRECOMPILE",
-                     False) or self.runner.model_config.enforce_eager:
+        if envs.SKIP_JAX_PRECOMPILE or self.runner.model_config.enforce_eager:
             return
         logger.info("Precompile all the subgraphs with possible input shapes.")
 
@@ -81,6 +82,8 @@ class CompilationManager:
                 self._precompile_backbone_with_inputs_embeds()
             if self.runner.scheduler_config.async_scheduling:
                 self._precompile_substitute_placeholder_token()
+            if not self.runner.is_last_rank:
+                return
             self._precompile_select_from_array()
             self._precompile_compute_logits()
             self._precompile_disagg_utils()
@@ -120,8 +123,15 @@ class CompilationManager:
                 num_tokens=num_tokens,
             )
 
-    def _precompile_backbone_helper(self, name, *, input_ids, positions,
-                                    inputs_embeds) -> None:
+    def _precompile_backbone_helper(self,
+                                    name,
+                                    *,
+                                    input_ids,
+                                    positions,
+                                    inputs_embeds,
+                                    intermediate_tensors=None,
+                                    is_first_rank=True,
+                                    is_last_rank=True) -> None:
         num_tokens = None
         if input_ids is not None:
             num_tokens = input_ids.shape[0]
@@ -135,12 +145,6 @@ class CompilationManager:
                 ShardingAxisName.ATTN_DATA, )) if dp_size > 1 else None
 
         # Keep existing pattern for complex array operations
-        block_tables = self.runner.block_table_cpu[:self.runner.max_num_reqs]
-        block_tables = block_tables.reshape(-1)
-        block_tables = device_array(self.runner.mesh,
-                                    block_tables,
-                                    sharding=dp_sharding)
-
         seq_lens = self._create_dummy_tensor((self.runner.max_num_reqs, ),
                                              jnp.int32, dp_sharding)
         query_start_loc = self._create_dummy_tensor(
@@ -152,26 +156,49 @@ class CompilationManager:
                                             request_distribution,
                                             sharding=dp_sharding)
 
-        attention_metadata = AttentionMetadata(
-            input_positions=positions,
-            block_tables=block_tables,
-            seq_lens=seq_lens,
-            query_start_loc=query_start_loc,
-            request_distribution=request_distribution,
-        )
+        attention_metadata_per_layer: Dict[str, AttentionMetadata] = {}
+        uniform_attention_metadata: AttentionMetadata = None
+        for kv_cache_gid, kv_cache_group in enumerate(
+                self.runner.kv_cache_config.kv_cache_groups):
+            block_tables = self.runner.block_tables_cpu[
+                kv_cache_gid][:self.runner.max_num_reqs]
+            block_tables = block_tables.reshape(-1)
+            block_tables = device_array(self.runner.mesh,
+                                        block_tables,
+                                        sharding=dp_sharding)
+
+            attention_metadata_gid = AttentionMetadata(
+                input_positions=positions,
+                block_tables=block_tables,
+                seq_lens=seq_lens,
+                query_start_loc=query_start_loc,
+                request_distribution=request_distribution,
+            )
+            if not self.runner.use_hybrid_kvcache:
+                # all layers share the same attention metadata
+                uniform_attention_metadata = attention_metadata_gid
+            else:
+                for layer_name in kv_cache_group.layer_names:
+                    attention_metadata_per_layer[
+                        layer_name] = attention_metadata_gid
 
         def model_fn_wrapper(
             state,
             kv_caches,
             input_ids,
             attention_metadata,
+            positions,
             inputs_embeds,
             layer_name_to_kvcache_index,
             lora_metadata,
+            intermediate_tensors,
+            is_first_rank,
+            is_last_rank,
         ):
             kv_caches, hidden_states, _ = self.runner.model_fn(
                 state, kv_caches, input_ids, attention_metadata, inputs_embeds,
-                layer_name_to_kvcache_index, lora_metadata)
+                positions, layer_name_to_kvcache_index, lora_metadata,
+                intermediate_tensors, is_first_rank, is_last_rank)
             self.runner.kv_caches = kv_caches
             return hidden_states
 
@@ -179,6 +206,10 @@ class CompilationManager:
                 self.runner.lora_config, np.array([num_tokens],
                                                   dtype=np.int32)):
             lora_metadata = self.runner.lora_utils.extract_lora_metadata()
+            if self.runner.use_hybrid_kvcache:
+                attention_metadata = attention_metadata_per_layer
+            else:
+                attention_metadata = uniform_attention_metadata
             self._run_compilation(
                 name,
                 model_fn_wrapper,
@@ -186,9 +217,13 @@ class CompilationManager:
                 self.runner.kv_caches,
                 input_ids,
                 attention_metadata,
+                positions,
                 inputs_embeds,
                 tuple(self.runner.layer_name_to_kvcache_index.items()),
                 lora_metadata,
+                intermediate_tensors,
+                is_first_rank,
+                is_last_rank,
                 num_tokens=num_tokens,
             )
 
@@ -239,6 +274,7 @@ class CompilationManager:
                 )
 
     def _precompile_backbone_text_only(self) -> None:
+        hidden_size = self.runner.model_config.get_hidden_size()
         for num_tokens in self.runner.num_tokens_paddings:
             dp_sharding = NamedSharding(
                 self.runner.mesh, PartitionSpec(ShardingAxisName.ATTN_DATA, )
@@ -248,10 +284,28 @@ class CompilationManager:
                                                   dp_sharding)
             positions = self._create_dummy_tensor((num_tokens, ), jnp.int32,
                                                   dp_sharding)
-            self._precompile_backbone_helper("backbone",
-                                             input_ids=input_ids,
-                                             positions=positions,
-                                             inputs_embeds=None)
+            is_first_rank = self.runner.is_first_rank
+            is_last_rank = self.runner.is_last_rank
+            if is_first_rank:
+                intermediate_tensors = None
+            else:
+                hidden_states = self._create_dummy_tensor(
+                    (num_tokens, hidden_size), jnp.bfloat16)
+                residual = self._create_dummy_tensor((num_tokens, hidden_size),
+                                                     jnp.bfloat16)
+                intermediate_tensors = JaxIntermediateTensors(
+                    tensors={
+                        "hidden_states": hidden_states,
+                        "residual": residual
+                    })
+            self._precompile_backbone_helper(
+                f"worker{self.runner.rank} backbone",
+                input_ids=input_ids,
+                positions=positions,
+                inputs_embeds=None,
+                intermediate_tensors=intermediate_tensors,
+                is_first_rank=is_first_rank,
+                is_last_rank=is_last_rank)
 
     def _precompile_backbone_with_inputs_embeds(self) -> None:
         hidden_size = self.runner.model_config.get_hidden_size()
@@ -265,10 +319,28 @@ class CompilationManager:
             else:
                 positions = self._create_dummy_tensor((num_tokens, ),
                                                       jnp.int32)
-            self._precompile_backbone_helper("backbone with embeds",
-                                             input_ids=None,
-                                             positions=positions,
-                                             inputs_embeds=inputs_embeds)
+            is_first_rank = self.runner.is_first_rank
+            is_last_rank = self.runner.is_last_rank
+            if not is_first_rank:
+                hidden_states = self._create_dummy_tensor(
+                    (num_tokens, hidden_size), jnp.bfloat16)
+                residual = self._create_dummy_tensor((num_tokens, hidden_size),
+                                                     jnp.bfloat16)
+                intermediate_tensors = JaxIntermediateTensors(
+                    tensors={
+                        "hidden_states": hidden_states,
+                        "residual": residual
+                    })
+            else:
+                intermediate_tensors = None
+            self._precompile_backbone_helper(
+                f"worker{self.runner.rank} backbone with embeds",
+                input_ids=None,
+                positions=positions,
+                inputs_embeds=inputs_embeds,
+                intermediate_tensors=intermediate_tensors,
+                is_first_rank=is_first_rank,
+                is_last_rank=is_last_rank)
 
     def _precompile_select_from_array_helper(
         self,
@@ -332,20 +404,23 @@ class CompilationManager:
             index_paddings = self.runner.num_reqs_paddings
         dp_sharding = NamedSharding(self.runner.mesh,
                                     PartitionSpec(ShardingAxisName.ATTN_DATA))
+        hidden_states_sharding = NamedSharding(
+            self.runner.mesh, PartitionSpec(ShardingAxisName.ATTN_DATA, None))
         dp_size = self.runner.vllm_config.sharding_config.total_dp_size
         self._precompile_select_from_array_helper(
-            name="select all logits",
+            name=f"worker{self.runner.rank} select all logits",
             source_paddings=self.runner.num_tokens_paddings,
             indices_paddings=index_paddings,
             hidden_dim=hsize,
-            input_sharding=dp_sharding,
+            input_sharding=hidden_states_sharding,
             indices_sharding=dp_sharding if dp_size > 1 else None,
         )
 
         if self.runner.speculative_config:
             vocab_size = self.runner.model_config.get_vocab_size()
             self._precompile_select_from_array_helper(
-                name="select bonus tokens for spec decoding",
+                name=
+                f"worker{self.runner.rank} select bonus tokens for spec decoding",
                 source_paddings=self.runner.num_logits_paddings,
                 indices_paddings=self.runner.num_reqs_paddings,
                 hidden_dim=vocab_size,
@@ -353,7 +428,8 @@ class CompilationManager:
                                              PartitionSpec(None, "model")),
             )
             self._precompile_select_from_array_helper(
-                name="select target tokens for spec decoding",
+                name=
+                f"worker{self.runner.rank} select target tokens for spec decoding",
                 source_paddings=self.runner.num_logits_paddings,
                 indices_paddings=self.runner.num_logits_paddings,
                 hidden_dim=vocab_size,
@@ -376,7 +452,7 @@ class CompilationManager:
                     np.array([num_reqs], dtype=np.int32)):
                 lora_metadata = self.runner.lora_utils.extract_lora_metadata()
                 self._run_compilation(
-                    "compute_logits",
+                    f"worker{self.runner.rank} compute_logits",
                     self.runner.compute_logits_fn,
                     self.runner.state,
                     hidden_states,
@@ -418,7 +494,7 @@ class CompilationManager:
                     do_sampling=do_sampling,
                 )
                 self._run_compilation(
-                    "sample",
+                    f"worker{self.runner.rank} sample",
                     sample,
                     self.runner.rng_params_for_sampling,
                     self.runner.mesh,
@@ -459,7 +535,7 @@ class CompilationManager:
             logits = self._create_dummy_tensor((num_reqs, hsize), jnp.bfloat16)
             token_ids = self._create_dummy_tensor((num_reqs, ), jnp.int32)
             self._run_compilation(
-                "gather_logprobs",
+                f"worker{self.runner.rank} gather_logprobs",
                 self.runner._compute_and_gather_logprobs,
                 logits,
                 token_ids,
@@ -511,7 +587,7 @@ class CompilationManager:
                             do_sampling=do_sampling)
 
                     self._run_compilation(
-                        compilation_name,
+                        f"worker{self.runner.rank} {compilation_name}",
                         self.runner.rejection_sampler,
                         draft_token_ids,
                         num_draft_tokens,
@@ -528,7 +604,9 @@ class CompilationManager:
     def _precompile_eagle3_helpers(self) -> None:
         logger.info(
             "Compiling eagle3 jitted helpers with different input shapes.")
-        hidden_size = self.runner.model_config.get_hidden_size()
+        target_hidden_size = self.runner.model_config.get_hidden_size()
+        draft_hidden_size = self.runner.speculative_config.draft_model_config.get_hidden_size(
+        )
         dtype = self.runner.model_config.dtype
 
         num_kv_cache_groups = len(self.runner.kv_cache_config.kv_cache_groups)
@@ -575,10 +653,11 @@ class CompilationManager:
 
         for num_logits in self.runner.num_logits_paddings:
             hidden_states = self._create_dummy_tensor(
-                (num_logits, hidden_size), jnp.bfloat16)
+                (num_logits, draft_hidden_size), jnp.bfloat16)
             self._run_compilation(
                 "eagle3_get_draft_token_ids",
                 self.runner.drafter._get_draft_token_ids,
+                self.runner.drafter.state,
                 hidden_states,
                 num_logits=num_logits,
             )
@@ -586,8 +665,8 @@ class CompilationManager:
         input_ids_loop = self._create_dummy_tensor(
             (self.runner.max_num_reqs, ), jnp.int32,
             NamedSharding(self.runner.mesh, PartitionSpec()))
-        target_hidden_state_loop = self._create_dummy_tensor(
-            (self.runner.max_num_reqs, hidden_size), dtype,
+        draft_hidden_state_loop = self._create_dummy_tensor(
+            (self.runner.max_num_reqs, draft_hidden_size), dtype,
             NamedSharding(self.runner.mesh, PartitionSpec(None, None)))
         next_token_ids = self._create_dummy_tensor(
             (self.runner.max_num_reqs, ), jnp.int32)
@@ -595,9 +674,12 @@ class CompilationManager:
             (self.runner.max_num_reqs, ), jnp.int32)
         for num_tokens in self.runner.num_tokens_paddings:
             aux_hidden_states = [
-                self._create_dummy_tensor((num_tokens, hidden_size), dtype),
-                self._create_dummy_tensor((num_tokens, hidden_size), dtype),
-                self._create_dummy_tensor((num_tokens, hidden_size), dtype),
+                self._create_dummy_tensor((num_tokens, target_hidden_size),
+                                          dtype),
+                self._create_dummy_tensor((num_tokens, target_hidden_size),
+                                          dtype),
+                self._create_dummy_tensor((num_tokens, target_hidden_size),
+                                          dtype),
             ]
 
             positions = self._create_dummy_tensor((num_tokens, ), jnp.int32)
@@ -620,23 +702,23 @@ class CompilationManager:
                 num_reqs,
             ):
                 target_hidden_states, input_ids, last_token_indices, _ = self.runner.drafter._filter_token_and_prepare_initial_inputs(
-                    token_indices, query_start_loc, seq_lens, input_ids,
-                    aux_hidden_states, attention_metadata, next_token_ids,
-                    num_reqs)
+                    self.runner.drafter.state, token_indices, query_start_loc,
+                    seq_lens, input_ids, aux_hidden_states, attention_metadata,
+                    next_token_ids, num_reqs)
                 return target_hidden_states, input_ids, last_token_indices
 
             input_ids = self._create_dummy_tensor((num_tokens, ), jnp.int32)
             aux_hidden_states = [
                 self._create_dummy_tensor(
-                    (num_tokens, hidden_size), jnp.bfloat16,
+                    (num_tokens, target_hidden_size), jnp.bfloat16,
                     NamedSharding(self.runner.mesh, PartitionSpec(None,
                                                                   None))),
                 self._create_dummy_tensor(
-                    (num_tokens, hidden_size), jnp.bfloat16,
+                    (num_tokens, target_hidden_size), jnp.bfloat16,
                     NamedSharding(self.runner.mesh, PartitionSpec(None,
                                                                   None))),
                 self._create_dummy_tensor(
-                    (num_tokens, hidden_size), jnp.bfloat16,
+                    (num_tokens, target_hidden_size), jnp.bfloat16,
                     NamedSharding(self.runner.mesh, PartitionSpec(None,
                                                                   None))),
             ]
@@ -668,17 +750,17 @@ class CompilationManager:
                 state,
                 kv_caches,
                 input_ids,
-                target_hidden_states,
+                draft_hidden_states,
                 attention_metadata,
             ):
                 kv_caches, hidden_states, _ = self.runner.drafter.model_fn(
-                    state, kv_caches, input_ids, target_hidden_states,
+                    state, kv_caches, input_ids, draft_hidden_states,
                     attention_metadata)
                 self.runner.kv_caches = kv_caches
                 return hidden_states
 
-            target_hidden_states = self._create_dummy_tensor(
-                (num_tokens, hidden_size), dtype,
+            draft_hidden_states = self._create_dummy_tensor(
+                (num_tokens, draft_hidden_size), dtype,
                 NamedSharding(self.runner.mesh, PartitionSpec(None, "model")))
             input_ids = self._create_dummy_tensor(
                 (num_tokens, ), jnp.int32,
@@ -689,7 +771,7 @@ class CompilationManager:
                 self.runner.drafter.state,
                 self.runner.kv_caches,
                 input_ids,
-                target_hidden_states,
+                draft_hidden_states,
                 attention_metadata,
                 num_tokens=num_tokens,
             )
@@ -699,6 +781,7 @@ class CompilationManager:
             self._run_compilation(
                 "eagle3_prepare_hidden_states_and_input_ids",
                 self.runner.drafter._prepare_hidden_states_and_input_ids,
+                self.runner.drafter.state,
                 aux_hidden_states,
                 query_start_loc,
                 target_token_ids,
@@ -721,18 +804,19 @@ class CompilationManager:
                 self.runner.drafter.state,
                 self.runner.kv_caches,
                 input_ids_loop,
-                target_hidden_state_loop,
+                draft_hidden_state_loop,
                 attention_metadata,
                 num_tokens=num_tokens,
             )
 
             hidden_states = self._create_dummy_tensor(
-                (num_tokens, hidden_size), jnp.bfloat16,
+                (num_tokens, draft_hidden_size), jnp.bfloat16,
                 NamedSharding(self.runner.mesh, PartitionSpec(None, None)))
 
             self._run_compilation(
                 "eagle3_select_inputs_for_loop_speculation",
                 self.runner.drafter._select_inputs_for_loop_speculation,
+                self.runner.drafter.state,
                 positions,
                 hidden_states,
                 hidden_states,
@@ -743,6 +827,7 @@ class CompilationManager:
             self._run_compilation(
                 "eagle3_select_draft_token_ids",
                 self.runner.drafter._select_draft_token_ids,
+                self.runner.drafter.state,
                 hidden_states,
                 last_token_indices,
                 num_tokens=num_tokens,

tpu_inference/runner/kv_cache_manager.py

@@ -1,15 +1,16 @@
 import functools
-import math
 from typing import TYPE_CHECKING, Dict, List
 
 import jax
 import jax.numpy as jnp
+import numpy as np
 import vllm.envs as envs
 from jax.sharding import NamedSharding, PartitionSpec
 from torchax.ops.mappings import t2j_dtype
-from vllm.attention import Attention
 from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.config import get_layers_from_vllm_config
+from vllm.utils.math_utils import cdiv
 from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
                                         KVCacheSpec, MLAAttentionSpec,
                                         SlidingWindowSpec)
@@ -175,6 +176,11 @@ class KVCacheManager:
             )
             self.runner.input_batch = new_input_batch
             self.runner.persistent_batch_manager.input_batch = new_input_batch
+            self.runner.block_tables_cpu = [
+                np.zeros((self.runner.max_num_reqs,
+                          cdiv(self.runner.max_model_len, block_size)),
+                         dtype=np.int32) for block_size in block_sizes
+            ]
 
     def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None:
         self.maybe_reinitialize_input_batch(kv_cache_config)
@@ -190,7 +196,7 @@ class KVCacheManager:
             num_blocks = kv_cache_tensor.size // page_size_bytes
             dp_size = self.runner.vllm_config.sharding_config.total_dp_size
             # num_blocks must be a multiple of dp_size
-            num_blocks = math.ceil(num_blocks / dp_size) * dp_size
+            num_blocks = (num_blocks // dp_size) * dp_size
             # NOTE: we'll multiply the num_kv_heads by 2 in the function
             kv_cache = create_kv_caches(
                 num_blocks=num_blocks,
@@ -283,13 +289,8 @@ class KVCacheManager:
 
         def _update_layer(cache, slices):
             """The function to apply to each layer's cache and slices."""
-            reshaped_slices = slices.reshape(-1, 1, block_size,
-                                             *slices.shape[1:])
-            for (i, block_idx) in enumerate(block_numbers):
-                cache = jax.lax.dynamic_update_slice_in_dim(cache,
-                                                            reshaped_slices[i],
-                                                            block_idx,
-                                                            axis=0)
+            reshaped_slices = slices.reshape(-1, block_size, *slices.shape[1:])
+            cache.at[block_numbers].set(reshaped_slices)
             return cache
 
         return jax.tree.map(_update_layer, kv_caches, kv_cache_slices)
@@ -342,16 +343,12 @@ class KVCacheManager:
         """
         if block_ids == list(range(block_ids[0],
                                    block_ids[0] + len(block_ids))):
-            with runner_utils.LatencyTracker(
-                    "BatchedGatherKVSlices-for-blocks"):
-                batched_kv_cache_per_layer = self._jitted_gather_continuous_kv_cache(
-                    self.runner.kv_caches, block_ids[0], len(block_ids))
+            batched_kv_cache_per_layer = self._jitted_gather_continuous_kv_cache(
+                self.runner.kv_caches, block_ids[0], len(block_ids))
 
         else:
-            with runner_utils.LatencyTracker(
-                    "BatchedGatherKVSlices-for-blocks"):
-                batched_kv_cache_per_layer = self._jitted_gather_kv_cache(
-                    self.runner.kv_caches, jnp.array(block_ids))
+            batched_kv_cache_per_layer = self._jitted_gather_kv_cache(
+                self.runner.kv_caches, jnp.array(block_ids))
         return batched_kv_cache_per_layer
 
     def transfer_kv_cache(self,
@@ -440,6 +437,7 @@ class KVCacheManager:
                     kv_cache_slices,
                     start_block,
                 )
+                jax.block_until_ready(self.runner.kv_caches)
         else:
             with runner_utils.LatencyTracker(
                     f"JittedInsertKVCache-b{len(block_numbers)}"):
@@ -451,6 +449,7 @@ class KVCacheManager:
                     kv_cache_slices,
                     jnp.array(block_numbers),
                 )
+                jax.block_until_ready(self.runner.kv_caches)
 
         logger.debug(
             f"Updated kv cache entries cnt={len(self.runner.kv_caches)}")

tpu_inference/runner/persistent_batch_manager.py

@@ -14,12 +14,13 @@ class PersistentBatchManager:
     def __init__(self, requests: Dict[str, CachedRequestState],
                  input_batch: InputBatch, encoder_cache: Dict[str,
                                                               'jax.Array'],
-                 uses_mrope: bool, model_config):
+                 uses_mrope: bool, model_config, is_last_rank: bool):
         self.requests = requests
         self.input_batch = input_batch
         self.encoder_cache = encoder_cache
         self.uses_mrope = uses_mrope
         self.model_config = model_config
+        self.is_last_rank = is_last_rank
 
     def _reorder_batch(self, scheduler_output: "VllmSchedulerOutput") -> int:
         """ Reorder the sheduled requests to RPA kernel friendly distribution
@@ -179,9 +180,35 @@ class PersistentBatchManager:
             num_computed_tokens = req_data.num_computed_tokens[i]
             new_block_ids = req_data.new_block_ids[i]
             resumed_from_preemption = req_data.resumed_from_preemption[i]
+            num_output_tokens = req_data.num_output_tokens[i]
 
             # Update the cached states.
             req_state.num_computed_tokens = num_computed_tokens
+            req_index = self.input_batch.req_id_to_index.get(req_id)
+
+            if not self.is_last_rank:
+                # When using PP, the scheduler sends the sampled tokens back,
+                # because there's no direct communication between the first-
+                # stage worker and the last-stage worker.
+                new_token_ids = req_data.new_token_ids[i]
+                # Add the sampled token(s) from the previous step (if any).
+                # This doesn't include "unverified" tokens like spec tokens.
+                num_new_tokens = (num_computed_tokens + len(new_token_ids) -
+                                  req_state.num_tokens)
+                if num_new_tokens == 1:
+                    req_state.output_token_ids.append(new_token_ids[-1])
+                elif num_new_tokens > 0:
+                    req_state.output_token_ids.extend(
+                        new_token_ids[-num_new_tokens:])
+            elif num_output_tokens < len(req_state.output_token_ids):
+                del req_state.output_token_ids[num_output_tokens:]
+                if req_index is not None:
+                    end_idx = (self.input_batch.num_prompt_tokens[req_index] +
+                               num_output_tokens)
+                    self.input_batch.num_tokens[req_index] = end_idx
+                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
+
+            # Update the block IDs.
             if not resumed_from_preemption:
                 if new_block_ids is not None:
                     # Append the new blocks to the existing block IDs.
@@ -194,7 +221,6 @@ class PersistentBatchManager:
                 # Replace the existing block IDs with the new ones.
                 req_state.block_ids = new_block_ids
 
-            req_index = self.input_batch.req_id_to_index.get(req_id)
             if req_index is None:
                 # The request is not in the persistent batch.
                 # The request was either preempted and resumed later, or was not
@@ -209,6 +235,18 @@ class PersistentBatchManager:
                 self.input_batch.block_table.append_row(
                     new_block_ids, req_index)
 
+            # For the last rank, we don't need to update the token_ids_cpu
+            # because the sampled tokens are already cached.
+            if not self.is_last_rank:
+                start_token_index = num_computed_tokens
+                end_token_index = num_computed_tokens + len(new_token_ids)
+                self.input_batch.token_ids_cpu[
+                    req_index,
+                    start_token_index:end_token_index] = new_token_ids
+                self.input_batch.num_tokens_no_spec[
+                    req_index] = end_token_index
+                self.input_batch.num_tokens[req_index] = end_token_index
+
             # Add spec_token_ids to token_ids_cpu.
             spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
                 req_id, ())

tpu_inference/runner/structured_decoding_manager.py

@@ -61,11 +61,10 @@ class StructuredDecodingManager:
         self.runner.require_structured_out_cpu.fill(0)
 
         sorted_struct_requests = sorted(
-            grammar_output.structured_output_request_ids.items(),
-            key=lambda item: item[1])
+            grammar_output.structured_output_request_ids)
 
         cumulative_mask_idx = 0
-        for req_id, _ in sorted_struct_requests:
+        for req_id in sorted_struct_requests:
             if req_id not in self.runner.input_batch.req_id_to_index:
                 continue
             batch_index = self.runner.input_batch.req_id_to_index[req_id]