tpu-inference 0.11.1.dev202511220812__py3-none-any.whl → 0.12.0.dev20251213__py3-none-any.whl

tpu_inference/runner/kv_cache_manager.py

@@ -7,8 +7,8 @@ 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,
@@ -39,20 +39,30 @@ class KVCacheManager:
         # means this layer will perform attention using the keys and values
         # from the KV cache of `shared_kv_cache_layers[layer_name]`.
         self.shared_kv_cache_layers: dict[str, str] = {}
+        self.use_mla = self.runner.model_config.use_mla
 
     def get_kv_cache_spec(self):
         # TODO(xiang): this hack tricks engine core to init successfully
         block_size = self.runner.cache_config.block_size
-        use_mla = self.runner.model_config.use_mla
         kv_cache_spec: dict[str, KVCacheSpec] = {}
 
         # If use pure jax (MODEL_IMPL_TYPE=flax_nnx), we don't register
         # attention into compilation config.
         # Use FullAttentionSpec for each layer
         # TODO(pooyam): Is it possible to merge the logic for vllm and non-vllm models?
+        model_config = self.runner.model_config
+        if self.use_mla:
+            # Individually pad the RopE and latents
+            qk_rope_head_dim = getattr(model_config.hf_text_config,
+                                       "qk_rope_head_dim", 0)
+            padded_kv_lora_rank = common_utils.align_to(
+                model_config.hf_text_config.kv_lora_rank, 128)
+            padded_qk_rope_head_dim = common_utils.align_to(
+                qk_rope_head_dim, 128)
+            mla_head_size = padded_kv_lora_rank + padded_qk_rope_head_dim
+
         if len(self.runner.vllm_config.compilation_config.
                static_forward_context) == 0:
-            model_config = self.runner.model_config
             parallel_config = self.runner.parallel_config
             # Pad num_kv_heads to multiple of TP size.
             num_kv_heads = common_utils.get_padded_num_heads(
@@ -61,11 +71,11 @@ class KVCacheManager:
             head_size = common_utils.get_padded_head_dim(
                 model_config.get_head_size())
             for i in range(model_config.get_num_layers(parallel_config)):
-                if use_mla:
+                if self.use_mla:
                     kv_cache_spec[f"layer.{i}"] = MLAAttentionSpec(
                         block_size=block_size,
-                        num_kv_heads=num_kv_heads,
-                        head_size=head_size,
+                        num_kv_heads=1,
+                        head_size=mla_head_size,
                         dtype=self.runner.kv_cache_dtype,
                         cache_dtype_str=self.runner.vllm_config.cache_config.
                         cache_dtype)
@@ -83,14 +93,13 @@ class KVCacheManager:
                     self.runner.mesh.shape["model"])
                 head_size = common_utils.get_padded_head_dim(
                     hf_config.hidden_size // hf_config.num_attention_heads)
-
                 # Eagle3 has only 1 layer
                 for i in range(1):
-                    if use_mla:
-                        kv_cache_spec[f"layer.{i}"] = MLAAttentionSpec(
+                    if self.use_mla:
+                        kv_cache_spec[f"draft_layer.{i}"] = MLAAttentionSpec(
                             block_size=block_size,
-                            num_kv_heads=num_kv_heads,
-                            head_size=head_size,
+                            num_kv_heads=1,
+                            head_size=mla_head_size,
                             dtype=self.runner.kv_cache_dtype,
                             cache_dtype_str=self.runner.vllm_config.
                             cache_config.cache_dtype)
@@ -104,6 +113,7 @@ class KVCacheManager:
             # Else propagate attention modules from compilation config.
             layers = get_layers_from_vllm_config(self.runner.vllm_config,
                                                  Attention)
+            logger.warning(f"Compilation num_layers = {len(layers.items())}")
             for layer_name, attn_module in layers.items():
                 if (kv_tgt_layer :=
                         attn_module.kv_sharing_target_layer_name) is not None:
@@ -127,11 +137,11 @@ class KVCacheManager:
                                 attn_module.head_size),
                             dtype=self.runner.kv_cache_dtype,
                             sliding_window=attn_module.sliding_window)
-                    elif use_mla:
-                        kv_cache_spec[f"layer.{i}"] = MLAAttentionSpec(
+                    elif self.use_mla:
+                        kv_cache_spec[layer_name] = MLAAttentionSpec(
                             block_size=block_size,
-                            num_kv_heads=attn_module.num_kv_heads,
-                            head_size=attn_module.head_size,
+                            num_kv_heads=1,
+                            head_size=mla_head_size,
                             dtype=self.runner.kv_cache_dtype,
                             cache_dtype_str=self.runner.vllm_config.
                             cache_config.cache_dtype)
@@ -198,14 +208,20 @@ class KVCacheManager:
             # num_blocks must be a multiple of dp_size
             num_blocks = (num_blocks // dp_size) * dp_size
             # NOTE: we'll multiply the num_kv_heads by 2 in the function
+            if self.use_mla:
+                head_size = self.runner.model_config.hf_config.kv_lora_rank + \
+                    self.runner.model_config.hf_config.qk_rope_head_dim
+            else:
+                head_size = representative_spec.head_size
             kv_cache = create_kv_caches(
                 num_blocks=num_blocks,
                 block_size=representative_spec.block_size,
                 num_kv_heads=representative_spec.num_kv_heads,
-                head_size=representative_spec.head_size,
+                head_size=head_size,
                 mesh=self.runner.mesh,
                 layer_names=[f'kv_cache_tensor.{i}'],
                 cache_dtype=t2j_dtype(representative_spec.dtype),
+                use_mla=self.use_mla,
             )[0]
             kv_caches.append(kv_cache)
             num_blocks_list.append(num_blocks)
@@ -289,13 +305,8 @@ 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)
@@ -348,16 +359,12 @@ 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,
@@ -446,6 +453,7 @@ 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)}"):
@@ -457,6 +465,7 @@ 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]

tpu_inference/runner/tpu_runner.py

@@ -1,6 +1,6 @@
 import copy
 import functools
-import os
+import logging
 import random
 from contextlib import nullcontext
 from dataclasses import dataclass
@@ -10,17 +10,15 @@ 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 +33,7 @@ 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 +47,8 @@ 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 +65,12 @@ 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 +81,6 @@ 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 +235,9 @@ 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 +248,21 @@ 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 +276,7 @@ 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 +287,7 @@ 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 +356,7 @@ 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 +398,7 @@ 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
         ]
@@ -555,12 +540,12 @@ 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 +671,9 @@ 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 +751,6 @@ 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 +761,14 @@ 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(
@@ -822,18 +814,31 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
 
         tpu_sampling_metadata = TPUSupportedSamplingMetadata.from_input_batch(
             self.mesh, self.input_batch, padded_num_reqs, sharding=sharding)
+
+        # TODO(pooyam): Should we move this to `_prepare_inputs`?
+        if tpu_sampling_metadata.do_sampling:
+            self.rng_params_for_sampling, step_rng = jax.random.split(
+                self.rng_params_for_sampling)
+        else:
+            step_rng = self.rng_params_for_sampling
+
         if spec_decode_metadata is None:
             next_tokens = sample(
-                self.rng_params_for_sampling,
+                step_rng,
                 self.mesh,
                 logits,
                 tpu_sampling_metadata,
             )
         else:
+            if tpu_sampling_metadata.do_sampling:
+                bonus_rng, rejection_rng = jax.random.split(step_rng)
+            else:
+                bonus_rng = step_rng
+                rejection_rng = step_rng
             bonus_logits = self._select_from_array_fn(
                 logits, spec_decode_metadata.bonus_logits_indices)
             bonus_token_ids = sample(
-                self.rng_params_for_sampling,
+                bonus_rng,
                 self.mesh,
                 bonus_logits,
                 tpu_sampling_metadata,
@@ -847,7 +852,7 @@ class TPUModelRunner(KVConnectorModelRunnerMixin, LoRAModelRunnerMixin):
                 target_logits=target_logits,
                 bonus_token_ids=bonus_token_ids,
                 sampling_metadata=tpu_sampling_metadata,
-                key=self.rng_params_for_sampling,
+                key=rejection_rng,
             )
 
         if tpu_sampling_metadata.logprobs:
@@ -1332,7 +1337,14 @@ 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(
@@ -1391,7 +1403,7 @@ 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(
@@ -1706,3 +1718,34 @@ 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 = ""