tpu-inference 0.11.1.dev202511270815__py3-none-any.whl → 0.13.0rc2.post7__py3-none-any.whl

tpu_inference/runner/tpu_runner.py

@@ -1,6 +1,20 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import copy
 import functools
-import os
+import logging
 import random
 from contextlib import nullcontext
 from dataclasses import dataclass
@@ -10,17 +24,15 @@ import jax
 import jax.numpy as jnp
 import jaxtyping
 import numpy as np
-import torch
-import vllm.envs as envs
+import vllm.envs as vllm_envs
 from flax import nnx
 from jax.experimental import mesh_utils
 from jax.sharding import NamedSharding, PartitionSpec
-from torchax.ops.mappings import j2t_dtype
 from vllm.config import VllmConfig
+from vllm.distributed import get_pp_group
 from vllm.distributed.kv_transfer import (get_kv_transfer_group,
                                           has_kv_transfer_group)
 from vllm.forward_context import set_forward_context
-from vllm.sequence import IntermediateTensors
 from vllm.tasks import SupportedTask
 from vllm.utils.math_utils import cdiv
 from vllm.v1.core.sched.output import GrammarOutput
@@ -35,6 +47,7 @@ from vllm.v1.worker.kv_connector_model_runner_mixin import \
     KVConnectorModelRunnerMixin
 from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
 
+import tpu_inference.envs as envs
 from tpu_inference import utils as common_utils
 from tpu_inference.layers.common.attention_metadata import AttentionMetadata
 from tpu_inference.layers.common.sharding import (MESH_AXIS_NAMES,
@@ -48,6 +61,8 @@ from tpu_inference.layers.jax.sample.sampling_metadata import \
     TPUSupportedSamplingMetadata
 from tpu_inference.logger import init_logger
 from tpu_inference.models.common.model_loader import get_model
+from tpu_inference.models.jax.jax_intermediate_tensor import \
+    JaxIntermediateTensors
 from tpu_inference.models.jax.utils.weight_utils import (
     shard_put, transfer_state_with_mappings)
 from tpu_inference.runner import utils as runner_utils
@@ -64,10 +79,12 @@ from tpu_inference.runner.structured_decoding_manager import \
     StructuredDecodingManager
 from tpu_inference.spec_decode.jax.eagle3 import Eagle3Proposer
 from tpu_inference.utils import (device_array, make_optimized_mesh,
-                                 time_function)
+                                 time_function, to_jax_dtype, to_torch_dtype)
 
 logger = init_logger(__name__)
 
+logging.getLogger("torchax.tensor").setLevel(logging.ERROR)
+
 INVALID_TOKEN_ID = -1
 # Smallest output size
 MIN_NUM_SEQS = 8
@@ -78,17 +95,6 @@ DUMMY_METADATA = AttentionMetadata(
     request_distribution=[0, 0, 0],
 )
 
-TPU_STR_DTYPE_TO_TORCH_DTYPE = {
-    "half": torch.half,
-    "bfloat16": torch.bfloat16,
-    "float": torch.float,
-    "fp8": torch.float8_e4m3fn,
-    "fp8_e4m3": torch.float8_e4m3fn,
-    "fp8_e5m2": torch.float8_e5m2,
-    "int8": torch.int8,
-    "uint8": torch.uint8,
-}
-
 
 class AsyncTPUModelRunnerOutput(AsyncModelRunnerOutput):
     """Holds asynchronous model output specifically from a TPU runner.
@@ -243,6 +249,9 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
         self.maybe_forbid_compile = runner_utils.ForbidCompile(
         ) if envs.VLLM_XLA_CHECK_RECOMPILATION else nullcontext()
         self.dp_size = self.vllm_config.sharding_config.total_dp_size
+        self.rank = rank
+        self.is_first_rank = is_first_rank
+        self.is_last_rank = is_last_rank
 
         self._init_random()
         self._init_mesh()
@@ -253,31 +262,21 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
 
         # Delegate functions to specific manager classes.
         self.compilation_manager = CompilationManager(self)
-        self.speculative_decoding_manager = SpeculativeDecodingManager(self)
-        self.structured_decoding_manager = StructuredDecodingManager(self)
+        if self.is_last_rank:
+            self.speculative_decoding_manager = SpeculativeDecodingManager(
+                self)
+            self.structured_decoding_manager = StructuredDecodingManager(self)
         self.kv_cache_manager = KVCacheManager(self)
         self.mm_manager = MultiModalManager(self)
         self.persistent_batch_manager = PersistentBatchManager(
             self.requests, self.input_batch, self.encoder_cache,
-            self.uses_mrope, self.model_config)
+            self.uses_mrope, self.model_config, self.is_last_rank)
         self.lora_utils = LoraUtils(self)
 
-        cache_config = self.cache_config
-        if cache_config.cache_dtype == "auto":
-            model_dtype = self.dtype
-            if isinstance(model_dtype, str):
-                self.kv_cache_dtype = TPU_STR_DTYPE_TO_TORCH_DTYPE[model_dtype]
-            elif isinstance(getattr(model_dtype, 'dtype', None), jnp.dtype):
-                self.kv_cache_dtype = j2t_dtype(model_dtype.dtype)
-            elif isinstance(model_dtype, torch.dtype):
-                self.kv_cache_dtype = model_dtype
-            else:
-                raise ValueError(
-                    "KV cache is unsupported for model_dtype of %s",
-                    model_dtype)
-        else:
-            self.kv_cache_dtype = TPU_STR_DTYPE_TO_TORCH_DTYPE[
-                cache_config.cache_dtype]
+        cache_dtype = self.cache_config.cache_dtype
+        if cache_dtype == "auto":
+            cache_dtype = self.dtype
+        self.kv_cache_dtype = to_torch_dtype(cache_dtype)
 
         self._pre_async_results: AsyncPreResults | None = None
         self._substitute_placeholder_token_fn = _substitute_placeholder_token
@@ -291,7 +290,7 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
         self.rng_key = jax.random.key(self.model_config.seed)
 
     def _init_mesh(self) -> None:
-        if os.getenv("NEW_MODEL_DESIGN", False):
+        if envs.NEW_MODEL_DESIGN:
             self.mesh = self._create_new_model_mesh()
         else:
             # NOTE(wenxindongwork): The new MoE kernel expects a 2D mesh, so we need
@@ -302,7 +301,7 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
         logger.info(f"Init mesh | mesh={self.mesh}")
 
     def _create_new_model_mesh(self) -> jax.sharding.Mesh:
-        num_slices = int(os.environ.get('NUM_SLICES', 1))
+        num_slices = envs.NUM_SLICES
 
         logger.info(f"Creating new model mesh | devices={len(self.devices)}, "
                     f"num_slices={num_slices}")
@@ -371,7 +370,7 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
                                        devices=self.devices)
 
     def _init_phased_profiling(self) -> None:
-        self.phased_profiling_dir = os.getenv("PHASED_PROFILING_DIR", "")
+        self.phased_profiling_dir = envs.PHASED_PROFILING_DIR
         self.phase_based_profiler = None
         if self.phased_profiling_dir:
             self.phase_based_profiler = runner_utils.PhasedBasedProfiler(
@@ -413,7 +412,7 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
             min_token_size=max(16, self.dp_size),
             max_token_size=scheduler_config.max_num_batched_tokens *
             self.dp_size,
-            padding_gap=envs.VLLM_TPU_BUCKET_PADDING_GAP)
+            padding_gap=vllm_envs.VLLM_TPU_BUCKET_PADDING_GAP)
         self.num_tokens_paddings_per_dp = [
             padding // self.dp_size for padding in self.num_tokens_paddings
         ]
@@ -509,10 +508,10 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
         multimodal_fns = multimodal_fns or {}
         self.precompile_vision_encoder_fn = multimodal_fns.get(
             "precompile_vision_encoder_fn", None)
-        self.get_multimodal_embeddings_fn = multimodal_fns.get(
-            "get_multimodal_embeddings_fn", None)
-        self.get_input_embeddings_fn = multimodal_fns.get(
-            "get_input_embeddings_fn", None)
+        self.embed_multimodal_fn = multimodal_fns.get("embed_multimodal_fn",
+                                                      None)
+        self.embed_input_ids_fn = multimodal_fns.get("embed_input_ids_fn",
+                                                     None)
         self.get_mrope_input_positions_fn = multimodal_fns.get(
             "get_mrope_input_positions_fn", None)
 
@@ -524,7 +523,7 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
             jax.random.key(self.model_config.seed)).params()
         self.is_multimodal_model = (
             self.model_config.is_multimodal_model
-            and self.get_multimodal_embeddings_fn is not None and hasattr(
+            and self.embed_multimodal_fn is not None and hasattr(
                 self.model_config.hf_config, "architectures"
             )  #TODO: Remove Llama Guard 4 specific condition once the LG4 Vision portion is implemented
             and len(self.model_config.hf_config.architectures) >= 1
@@ -540,7 +539,10 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
     def get_kv_cache_spec(self):
         return self.kv_cache_manager.get_kv_cache_spec()
 
-    def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None:
+    def initialize_kv_cache(self,
+                            kv_cache_config: KVCacheConfig,
+                            topology_order_id: int = 0) -> None:
+        self.topology_order_id = topology_order_id
         self.kv_cache_config = kv_cache_config
         self.use_hybrid_kvcache = len(kv_cache_config.kv_cache_groups) > 1
         self.kv_caches = []
@@ -555,12 +557,12 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
     def execute_model(
         self,
         scheduler_output: "VllmSchedulerOutput",
-        intermediate_tensors: Optional[IntermediateTensors] = None,
-    ) -> ModelRunnerOutput | None:
+        intermediate_tensors: Optional[JaxIntermediateTensors] = None,
+    ) -> ModelRunnerOutput | JaxIntermediateTensors | None:
         if self.execute_model_state is not None:
             raise RuntimeError("State error: sample_tokens() must be called "
                                "after execute_model() returns None.")
-        _, output = self._execute_model(scheduler_output)
+        _, output = self._execute_model(scheduler_output, intermediate_tensors)
         return output
 
     def sample_tokens(
@@ -686,7 +688,9 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
     def _execute_model(
         self,
         scheduler_output: "VllmSchedulerOutput",
-    ) -> tuple[AttentionMetadata, ModelRunnerOutput | None]:
+        intermediate_tensors: Optional[JaxIntermediateTensors] = None,
+    ) -> tuple[AttentionMetadata, JaxIntermediateTensors | ModelRunnerOutput
+               | None]:
         self.persistent_batch_manager.update_states(
             scheduler_output, self.get_mrope_input_positions_fn)
         if not scheduler_output.total_num_scheduled_tokens:
@@ -764,7 +768,6 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
                     scheduler_output) as kv_connector_output:
                 # NOTE(Wenlong): It takes both `input_ids` and `inputs_embeds`,
                 # but one of them would be `None`
-
                 (self.kv_caches, hidden_states,
                  aux_hidden_states) = self.model_fn(
                      self.state,
@@ -775,8 +778,14 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
                      input_positions,
                      tuple(self.layer_name_to_kvcache_index.items()),
                      lora_metadata,
+                     intermediate_tensors,
+                     self.is_first_rank,
+                     self.is_last_rank,
                  )
-
+            if not get_pp_group().is_last_rank:
+                assert isinstance(hidden_states, JaxIntermediateTensors)
+                hidden_states.kv_connector_output = kv_connector_output
+                return attn_metadata, hidden_states
             hidden_states = self._select_from_array_fn(hidden_states,
                                                        logits_indices)
             logits = self.compute_logits_fn(
@@ -818,7 +827,7 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
         sharding = None
         if self.dp_size > 1:
             sharding = NamedSharding(self.mesh,
-                                     PartitionSpec(ShardingAxisName.ATTN_DATA))
+                                     PartitionSpec(ShardingAxisName.MLP_DATA))
 
         tpu_sampling_metadata = TPUSupportedSamplingMetadata.from_input_batch(
             self.mesh, self.input_batch, padded_num_reqs, sharding=sharding)
@@ -1345,7 +1354,14 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
         _request_distribution = []
         for dp_rank in range(dp_size):
             _num_reqs = num_req_per_dp_rank[dp_rank]
-            _request_distribution.append([0, 0, _num_reqs])
+            # The batch has been reordered by _reorder_batch so decode requests come first
+            # Count decode requests (those with num_scheduled_tokens == 1) in this DP rank
+            num_decode_in_dp_rank = 0
+            for req_id in req_ids_dp[dp_rank]:
+                if scheduler_output.num_scheduled_tokens[req_id] == 1:
+                    num_decode_in_dp_rank += 1
+            _request_distribution.append(
+                [num_decode_in_dp_rank, num_decode_in_dp_rank, _num_reqs])
         request_distribution = np.array(_request_distribution).ravel()
 
         use_spec_decode = len(
@@ -1374,7 +1390,8 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
             self.mesh,
             self.input_batch,
             padded_num_reqs,
-            sharding=data_parallel_attn_sharding,
+            sharding=NamedSharding(self.mesh,
+                                   PartitionSpec(ShardingAxisName.MLP_DATA)),
         )
         if self.uses_mrope:
             positions = mrope_positions
@@ -1404,7 +1421,7 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
                 block_tables[
                     req_offset:req_offset + _num_reqs, :self.
                     max_num_blocks_per_req] = self.input_batch.block_table[
-                        0].get_cpu_tensor()[req_indices_dp[dp_rank]]
+                        kv_cache_gid].get_cpu_tensor()[req_indices_dp[dp_rank]]
             # Convert block_tables to 1D on cpu.
             block_tables = block_tables.reshape(-1)
             block_tables = device_array(
@@ -1664,7 +1681,7 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
     def _get_input_ids_embeds(self, input_ids: jax.Array,
                               mm_embeds: list[jax.Array]):
         if self.is_multimodal_model:
-            inputs_embeds = self.get_input_embeddings_fn(
+            inputs_embeds = self.embed_input_ids_fn(
                 self.state,
                 input_ids,
                 mm_embeds,
@@ -1719,3 +1736,34 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
             mappings=mappings,
             transpose_keys=transpose_keys,
             shard=shard)
+
+    def get_intermediate_tensor_spec(self, num_tokens: int):
+        jax_dtype = to_jax_dtype(self.dtype)
+        num_padded_tokens = runner_utils.get_padded_token_len(
+            self.num_tokens_paddings, num_tokens)
+        sharding = NamedSharding(self.mesh, PartitionSpec())
+        hidden_size = self.model_config.get_hidden_size()
+        spec = jax.ShapeDtypeStruct(shape=(num_padded_tokens, hidden_size),
+                                    dtype=jax_dtype,
+                                    sharding=sharding)
+        tensor_spec = {"hidden_states": spec, "residual": spec}
+        return tensor_spec
+
+    def get_uuid_for_jax_transfer(self,
+                                  scheduler_output: "VllmSchedulerOutput",
+                                  rank: int, step: int) -> int:
+        '''
+        Get a uuid for jax.transfer, here we use the hash of
+        scheduler_output + counter_step + sender's rank
+        '''
+        scheduler_output_str = ""
+        if not scheduler_output.num_scheduled_tokens:
+            scheduler_output_str = "empty_batch"
+        else:
+            scheduler_output_str = str(
+                sorted(scheduler_output.num_scheduled_tokens.items()))
+        unique_str = f'{scheduler_output_str} {step} {rank}'
+        import hashlib
+        hasher = hashlib.sha1()
+        hasher.update(unique_str.encode('utf-8'))
+        return int.from_bytes(hasher.digest()[:8], 'big')

tpu_inference/runner/utils.py

@@ -15,6 +15,7 @@ import jax
 from jax._src.interpreters import pxla
 from vllm.v1.core.sched.output import SchedulerOutput as VllmSchedulerOutput
 
+from tpu_inference import envs
 from tpu_inference.logger import init_logger
 from tpu_inference.runner.input_batch import InputBatch
 
@@ -306,8 +307,7 @@ class PhasedBasedProfiler:
             InferencePhase.BALANCED: False
         }
         self.default_profiling_options = jax.profiler.ProfileOptions()
-        self.default_profiling_options.python_tracer_level = os.getenv(
-            "PYTHON_TRACER_LEVEL", 0)
+        self.default_profiling_options.python_tracer_level = envs.PYTHON_TRACER_LEVEL
 
         self.current_phase: str = ""
 

tpu_inference/spec_decode/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

tpu_inference/spec_decode/jax/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

tpu_inference/spec_decode/jax/eagle3.py

@@ -1,3 +1,16 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 """Implements the Eagle3 proposer for speculative decoding on JAX/TPU."""
 import functools
 from dataclasses import replace
@@ -6,6 +19,9 @@ from typing import Any, Optional
 import jax
 import jax.numpy as jnp
 import numpy as np
+from flax import nnx
+from jax import lax
+from jax.sharding import NamedSharding, PartitionSpec
 from vllm.config import VllmConfig
 
 from tpu_inference.layers.common.attention_metadata import AttentionMetadata
@@ -127,6 +143,17 @@ class Eagle3Proposer:
                                            max_num_blocks_per_req)
         new_block_tables = jnp.where(expanded_exceeds_mask, -1, block_tables)
 
+        positions = lax.with_sharding_constraint(
+            positions, NamedSharding(self.mesh, PartitionSpec(None, )))
+        clamped_positions = lax.with_sharding_constraint(
+            clamped_positions, NamedSharding(self.mesh, PartitionSpec(None, )))
+        new_seq_lens = lax.with_sharding_constraint(
+            new_seq_lens, NamedSharding(self.mesh, PartitionSpec(None, )))
+        query_start_loc = lax.with_sharding_constraint(
+            query_start_loc, NamedSharding(self.mesh, PartitionSpec()))
+        new_block_tables = lax.with_sharding_constraint(
+            new_block_tables, NamedSharding(self.mesh, PartitionSpec(None, )))
+
         return positions, clamped_positions, new_seq_lens, query_start_loc, new_block_tables
 
     @functools.partial(jax.jit, static_argnums=(0, ))
@@ -138,6 +165,7 @@ class Eagle3Proposer:
     @functools.partial(jax.jit, static_argnums=(0, ))
     def _prepare_hidden_states_and_input_ids(
         self,
+        state: nnx.State,
         aux_hidden_states: tuple[jax.Array, ...],
         query_start_loc: jax.Array,
         target_token_ids: jax.Array,
@@ -146,7 +174,7 @@ class Eagle3Proposer:
     ) -> tuple[jax.Array, jax.Array, jax.Array]:
         target_hidden_states = jnp.concatenate(aux_hidden_states, axis=-1)
         target_hidden_states = self.combine_hidden_states_fn(
-            self.state, target_hidden_states)
+            state, target_hidden_states)
 
         input_ids, last_token_indices = self._prepare_input_ids(
             query_start_loc, target_token_ids, next_token_ids, num_reqs)
@@ -193,8 +221,8 @@ class Eagle3Proposer:
                                     block_tables=device_array(
                                         self.mesh, block_tables))
             target_hidden_states, input_ids, last_token_indices = self._prepare_hidden_states_and_input_ids(
-                aux_hidden_states, attn_metadata.query_start_loc, input_ids,
-                next_token_ids, num_reqs)
+                self.state, aux_hidden_states, attn_metadata.query_start_loc,
+                input_ids, next_token_ids, num_reqs)
             return target_hidden_states, input_ids, last_token_indices, attn_metadata
 
         # Host copies from the metadata prepared by the runner.
@@ -258,12 +286,13 @@ class Eagle3Proposer:
 
         attn_metadata = replace(attn_metadata, block_tables=block_tables)
         return self._filter_token_and_prepare_initial_inputs(
-            token_indices, query_start_loc, seq_lens, input_ids,
+            self.state, token_indices, query_start_loc, seq_lens, input_ids,
             aux_hidden_states, attn_metadata, next_token_ids, num_reqs)
 
     @functools.partial(jax.jit, static_argnums=(0, ))
     def _filter_token_and_prepare_initial_inputs(
         self,
+        state: nnx.State,
         token_indices: jax.Array,
         query_start_loc: jax.Array,
         seq_lens: jax.Array,
@@ -291,35 +320,51 @@ class Eagle3Proposer:
         )
 
         target_hidden_states, input_ids, last_token_indices = self._prepare_hidden_states_and_input_ids(
-            [h[token_indices] for h in aux_hidden_states], query_start_loc,
-            target_token_ids, next_token_ids, num_reqs)
+            state, [h[token_indices] for h in aux_hidden_states],
+            query_start_loc, target_token_ids, next_token_ids, num_reqs)
 
         return target_hidden_states, input_ids, last_token_indices, attn_metadata
 
     @functools.partial(jax.jit, static_argnums=(0, ))
     def _select_draft_token_ids(
         self,
+        state: nnx.State,
         hidden_states: jax.Array,
         last_token_indices: jax.Array,
     ) -> jax.Array:
         sample_hidden_states = hidden_states[last_token_indices]
-        return self._get_draft_token_ids(sample_hidden_states)
+        sample_hidden_states = lax.with_sharding_constraint(
+            sample_hidden_states,
+            NamedSharding(self.mesh, PartitionSpec(None, None)))
+        return self._get_draft_token_ids(state, sample_hidden_states)
 
     @functools.partial(jax.jit, static_argnums=(0, ))
-    def _get_draft_token_ids(self, hidden_states: jax.Array) -> jax.Array:
+    def _get_draft_token_ids(self, state: nnx.State,
+                             hidden_states: jax.Array) -> jax.Array:
         lora_metadata = None
-        logits = self.compute_logits_fn(self.state, hidden_states,
-                                        lora_metadata)
-        return jnp.argmax(logits, axis=-1)
+        logits = self.compute_logits_fn(state, hidden_states, lora_metadata)
+        draft_token_ids = jnp.argmax(logits, axis=-1)
+        return lax.with_sharding_constraint(
+            draft_token_ids, NamedSharding(self.mesh, PartitionSpec()))
 
     @functools.partial(jax.jit, static_argnums=(0, ))
     def _select_inputs_for_loop_speculation(
-            self, positions: jax.Array, residual: jax.Array,
+            self, state: nnx.State, positions: jax.Array, residual: jax.Array,
             hidden_states: jax.Array,
             last_token_indices: jax.Array) -> tuple[jax.Array, jax.Array]:
-        return positions[last_token_indices], residual[
-            last_token_indices], self._select_draft_token_ids(
-                hidden_states, last_token_indices)
+        positions = positions[last_token_indices]
+        residual = residual[last_token_indices]
+        draft_token_ids = self._select_draft_token_ids(state, hidden_states,
+                                                       last_token_indices)
+
+        positions = lax.with_sharding_constraint(
+            positions, NamedSharding(self.mesh, PartitionSpec(None, )))
+        residual = lax.with_sharding_constraint(
+            residual, NamedSharding(self.mesh, PartitionSpec(None, None)))
+        draft_token_ids = lax.with_sharding_constraint(
+            draft_token_ids, NamedSharding(self.mesh, PartitionSpec()))
+
+        return positions, residual, draft_token_ids
 
     def propose(
         self,
@@ -346,11 +391,11 @@ class Eagle3Proposer:
 
         if self.num_speculative_tokens == 1:
             return kv_caches, self._select_draft_token_ids(
-                hidden_states, last_token_indices)
+                self.state, hidden_states, last_token_indices)
 
         positions, hidden_states, draft_token_ids = self._select_inputs_for_loop_speculation(
-            attn_metadata.input_positions, residual[0], hidden_states,
-            last_token_indices)
+            self.state, attn_metadata.input_positions, residual[0],
+            hidden_states, last_token_indices)
 
         draft_token_ids_list = [draft_token_ids]
 
@@ -375,7 +420,8 @@ class Eagle3Proposer:
                 attn_metadata,
             )
             hidden_states = residual[0]
-            draft_token_ids = self._get_draft_token_ids(new_hidden_states)
+            draft_token_ids = self._get_draft_token_ids(
+                self.state, new_hidden_states)
             draft_token_ids_list.append(draft_token_ids)
 
         # [batch_size, num_speculative_tokens]

tpu_inference/tpu_info.py

@@ -1,3 +1,17 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import glob
 import os