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

tpu_inference/runner/multimodal_manager.py

@@ -0,0 +1,217 @@
+from typing import TYPE_CHECKING
+
+import jax
+import jax.numpy as jnp
+from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
+from vllm.multimodal.inputs import MultiModalKwargsItem, PlaceholderRange
+from vllm.multimodal.utils import group_mm_kwargs_by_modality
+from vllm.v1.core.sched.output import SchedulerOutput as VllmSchedulerOutput
+from vllm.v1.worker.utils import (gather_mm_placeholders,
+                                  scatter_mm_placeholders)
+
+from tpu_inference.models.jax.utils.multi_modal_utils import (
+    flatten_embeddings, sanity_check_mm_encoder_outputs)
+
+if TYPE_CHECKING:
+    from tpu_inference.runner.tpu_runner import TPUModelRunner
+
+
+class MultiModalManager:
+
+    def __init__(self, runner: "TPUModelRunner"):
+        self.runner = runner
+
+    def calc_mrope_positions(self, scheduler_output: "VllmSchedulerOutput"):
+        mrope_pos_ptr = 0
+        for index, req_id in enumerate(self.runner.input_batch.req_ids):
+            req = self.runner.requests[req_id]
+            assert req.mrope_positions is not None
+
+            num_computed_tokens = \
+                self.runner.input_batch.num_computed_tokens_cpu[index]
+            num_scheduled_tokens = \
+                scheduler_output.num_scheduled_tokens[req_id]
+            num_prompt_tokens = len(req.prompt_token_ids)
+
+            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
+                prompt_part_len = max(0,
+                                      num_prompt_tokens - num_computed_tokens)
+                completion_part_len = max(
+                    0, num_scheduled_tokens - prompt_part_len)
+            else:
+                prompt_part_len = num_scheduled_tokens
+                completion_part_len = 0
+
+            assert num_scheduled_tokens == prompt_part_len + completion_part_len
+
+            if prompt_part_len > 0:
+                # prompt's mrope_positions are pre-computed
+                dst_start = mrope_pos_ptr
+                dst_end = mrope_pos_ptr + prompt_part_len
+                src_start = num_computed_tokens
+                src_end = num_computed_tokens + prompt_part_len
+
+                self.runner.mrope_positions_cpu[:, dst_start:dst_end] = \
+                    req.mrope_positions[:,src_start:src_end]
+
+                mrope_pos_ptr += prompt_part_len
+
+            if completion_part_len > 0:
+                # compute completion's mrope_positions on-the-fly
+                dst_start = mrope_pos_ptr
+                dst_end = mrope_pos_ptr + completion_part_len
+
+                MRotaryEmbedding.get_next_input_positions_tensor(
+                    out=self.runner.mrope_positions_cpu,
+                    out_offset=dst_start,
+                    mrope_position_delta=req.mrope_position_delta,
+                    context_len=num_computed_tokens + prompt_part_len,
+                    num_new_tokens=completion_part_len,
+                )
+
+                mrope_pos_ptr += completion_part_len
+
+    def execute_mm_encoder(self, scheduler_output: "VllmSchedulerOutput"):
+        import torch
+        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
+        if not scheduled_encoder_inputs:
+            return
+
+        # Batch the multi-modal inputs.
+        mm_kwargs = list[MultiModalKwargsItem]()
+        # List of tuple (mm_hash, pos_info)
+        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
+        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
+            req_state = self.runner.requests[req_id]
+            for mm_input_id in encoder_input_ids:
+                mm_feature = req_state.mm_features[mm_input_id]
+                mm_hash = mm_feature.identifier
+                mm_kwargs.append(mm_feature.data)
+                mm_hashes_pos.append((mm_hash, mm_feature.mm_position))
+
+        # Batch mm inputs as much as we can: if a request in the batch has
+        # multiple modalities or a different modality than the previous one,
+        # we process it separately to preserve item order.
+        # FIXME(ywang96): This is a hacky way to deal with multiple modalities
+        # in the same batch while still being able to benefit from batching
+        # multimodal inputs. The proper solution should be reordering the
+        # encoder outputs.
+        encoder_outputs = []
+        for _, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
+                mm_kwargs, merge_by_field_config=False):
+            batched_mm_inputs = mm_kwargs_group
+            # Convert torch tensors to numpy arrays that JAX can handle.
+            if "pixel_values" in batched_mm_inputs and isinstance(
+                    batched_mm_inputs["pixel_values"], list):
+                batched_mm_inputs["pixel_values"] = torch.cat(
+                    batched_mm_inputs["pixel_values"], dim=0)
+
+            image_grid_thw = ()
+            for key, value in batched_mm_inputs.items():
+                if isinstance(value, torch.Tensor):
+                    if key == 'image_grid_thw':
+                        # change it to tuple of tuples to make it hashable for JIT
+
+                        # Shape: (B, N, 3) -> (B*N, 3) -> tuple of tuples
+                        grid_thw_tensor = batched_mm_inputs[key]
+                        grid_thw_reshaped = grid_thw_tensor.reshape(-1, 3)
+                        image_grid_thw = tuple(
+                            tuple(row) for row in grid_thw_reshaped.tolist())
+
+                        continue
+
+                    if value.dtype == torch.bfloat16:
+                        batched_mm_inputs[key] = value.to(
+                            torch.float32).numpy().astype(jnp.bfloat16)
+                    else:
+                        batched_mm_inputs[key] = value.numpy()
+            batched_mm_inputs.pop('image_grid_thw')
+
+            # Run the encoder.
+            # `curr_group_outputs` is either of the following:
+            # 1. A tensor of shape (num_items, feature_size, hidden_size)
+            # in case feature_size is fixed across all multimodal items.
+            # 2. A list or tuple (length: num_items) of tensors, each of shape
+            # (feature_size, hidden_size) in case the feature size is dynamic
+            # depending on the input multimodal items.
+            curr_group_outputs = self.runner.get_multimodal_embeddings_fn(
+                self.runner.state, image_grid_thw, **batched_mm_inputs)
+
+            sanity_check_mm_encoder_outputs(
+                curr_group_outputs,
+                expected_num_items=num_items,
+            )
+
+            for output in curr_group_outputs:
+                encoder_outputs.append(output)
+
+        # Cache the encoder outputs.
+        for (mm_hash, pos_info), output in zip(
+                mm_hashes_pos,
+                encoder_outputs,
+        ):
+            if req_id not in self.runner.encoder_cache:
+                self.runner.encoder_cache[req_id] = {}
+
+            self.runner.encoder_cache[mm_hash] = scatter_mm_placeholders(
+                output,
+                is_embed=pos_info.is_embed,
+            )
+
+    def gather_mm_embeddings(self, scheduler_output: "VllmSchedulerOutput",
+                             target_pad_len: int) -> list[jax.Array]:
+        mm_embeds: list[jax.Array] = []
+        for req_id in self.runner.input_batch.req_ids:
+            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[
+                req_id]
+            req_state = self.runner.requests[req_id]
+            num_computed_tokens = req_state.num_computed_tokens
+            mm_features = req_state.mm_features
+            for _, mm_feature in enumerate(mm_features):
+                pos_info = mm_feature.mm_position
+                start_pos = pos_info.offset
+                num_encoder_tokens = pos_info.length
+
+                # The encoder output is needed if the two ranges overlap:
+                # [num_computed_tokens,
+                #  num_computed_tokens + num_scheduled_tokens) and
+                # [start_pos, start_pos + num_encoder_tokens)
+                if start_pos >= num_computed_tokens + num_scheduled_tokens:
+                    # The encoder output is not needed in this step.
+                    break
+                if start_pos + num_encoder_tokens <= num_computed_tokens:
+                    # The encoder output is already processed and stored
+                    # in the decoder's KV cache.
+                    continue
+
+                start_idx = max(num_computed_tokens - start_pos, 0)
+                end_idx = min(
+                    num_computed_tokens - start_pos + num_scheduled_tokens,
+                    num_encoder_tokens)
+                assert start_idx < end_idx
+                mm_hash = mm_feature.identifier
+                encoder_output = self.runner.encoder_cache.get(mm_hash, None)
+                assert encoder_output is not None,\
+                      f"Encoder cache miss for {mm_hash}."
+                encoder_output = self.runner.encoder_cache[mm_hash]
+
+                if (is_embed := pos_info.is_embed) is not None:
+                    is_embed = is_embed[start_idx:end_idx]
+
+                mm_embeds_item = gather_mm_placeholders(
+                    encoder_output[start_idx:end_idx],
+                    is_embed=is_embed,
+                )
+                mm_embeds.append(mm_embeds_item)
+        if not mm_embeds:
+            return None
+        flattened_embeds = flatten_embeddings(mm_embeds)
+        if flattened_embeds.shape[0] == 0:
+            return None
+
+        padding = jnp.zeros((target_pad_len - flattened_embeds.shape[0],
+                             flattened_embeds.shape[1]),
+                            dtype=flattened_embeds.dtype)
+        flattened_embeds = jnp.concatenate([flattened_embeds, padding], axis=0)
+
+        return flattened_embeds

tpu_inference/runner/persistent_batch_manager.py

@@ -0,0 +1,244 @@
+from typing import Dict
+
+import jax
+from vllm.v1.core.sched.output import SchedulerOutput as VllmSchedulerOutput
+
+from tpu_inference.logger import init_logger
+from tpu_inference.runner.input_batch import CachedRequestState, InputBatch
+
+logger = init_logger(__name__)
+
+
+class PersistentBatchManager:
+
+    def __init__(self, requests: Dict[str, CachedRequestState],
+                 input_batch: InputBatch, encoder_cache: Dict[str,
+                                                              'jax.Array'],
+                 uses_mrope: bool, model_config):
+        self.requests = requests
+        self.input_batch = input_batch
+        self.encoder_cache = encoder_cache
+        self.uses_mrope = uses_mrope
+        self.model_config = model_config
+
+    def _reorder_batch(self, scheduler_output: "VllmSchedulerOutput") -> int:
+        """ Reorder the sheduled requests to RPA kernel friendly distribution
+        (decode_only, fixed_chunked_prefill_only, mixed) and set the request
+        distribution accordingly.
+
+        Returns:
+            The number of swaps in requests.
+        """
+        # Note(jevinjiang): currently we only consider decode_only.
+        num_reqs = self.input_batch.num_reqs
+        swap_cnt = 0
+        if num_reqs <= 0:
+            return swap_cnt
+        # Use two-pointer approach to reorder the decode requests to front.
+        i, j = 0, num_reqs - 1
+        while i < j:
+            i_req_id = self.input_batch.req_ids[i]
+            j_req_id = self.input_batch.req_ids[j]
+
+            if scheduler_output.num_scheduled_tokens[i_req_id] == 1:
+                # i is a decode request, move to the next one.
+                i += 1
+            elif scheduler_output.num_scheduled_tokens[j_req_id] > 1:
+                # j is a prefill request, move to the previous one.
+                j -= 1
+            else:
+                # Swap i and j.
+                self.input_batch.swap_states(i, j)
+                i += 1
+                j -= 1
+                swap_cnt += 1
+
+        num_decode = i + int(scheduler_output.num_scheduled_tokens[
+            self.input_batch.req_ids[i]] == 1)
+
+        self.input_batch.request_distribution = [
+            num_decode, num_decode, num_reqs
+        ]
+
+        return swap_cnt
+
+    def update_states(self, scheduler_output: "VllmSchedulerOutput",
+                      get_mrope_input_positions_fn) -> bool:
+        """Update the cached states and the persistent batch with the scheduler
+        output.
+
+        The updated states are used by the `_prepare_inputs` function to create
+        the input TPU tensors for the model.
+
+        Returns:
+            True if there is a new/resumed/paused/finished request.
+            If False, we can skip copying SamplingMetadata to the TPU.
+        """
+        # Remove finished requests from the cached states.
+        for req_id in scheduler_output.finished_req_ids:
+            self.requests.pop(req_id, None)
+
+        # Remove the finished requests from the persistent batch.
+        # NOTE(woosuk): There could be an edge case where finished_req_ids and
+        # scheduled_req_ids overlap. This happens when a request is aborted and
+        # then resubmitted with the same ID. In this case, we treat them as two
+        # distinct requests - clearing the cached states for the first request
+        # and handling the second as a new request.
+        removed_req_indices: list[int] = []
+        for req_id in scheduler_output.finished_req_ids:
+            req_index = self.input_batch.remove_request(req_id)
+            if req_index is not None:
+                removed_req_indices.append(req_index)
+
+        # Free the cached encoder outputs.
+        for mm_hash in scheduler_output.free_encoder_mm_hashes:
+            self.encoder_cache.pop(mm_hash, None)
+
+        # Remove the unscheduled requests from the persistent batch.
+        # NOTE(woosuk): The unscheduled requests are either preempted requests
+        # or running requests that are not scheduled in this step. We remove
+        # them from the persistent batch but keep their cached states since
+        # they will be scheduled again sometime in the future.
+        scheduled_req_ids = scheduler_output.num_scheduled_tokens.keys()
+        cached_req_ids = self.input_batch.req_id_to_index.keys()
+        unscheduled_req_ids = cached_req_ids - scheduled_req_ids
+        # NOTE(woosuk): The persistent batch optimization assumes that
+        # consecutive batches contain mostly the same requests. If batches
+        # have low request overlap (e.g., alternating between two distinct
+        # sets of requests), this optimization becomes very inefficient.
+        for req_id in unscheduled_req_ids:
+            req_index = self.input_batch.remove_request(req_id)
+            assert req_index is not None
+            removed_req_indices.append(req_index)
+
+        req_ids_to_add: list[str] = []
+        # Add new requests to the cached states.
+        for new_req_data in scheduler_output.scheduled_new_reqs:
+            req_id = new_req_data.req_id
+            sampling_params = new_req_data.sampling_params
+
+            self.requests[req_id] = CachedRequestState(
+                req_id=req_id,
+                prompt_token_ids=new_req_data.prompt_token_ids,
+                mm_features=new_req_data.mm_features,
+                sampling_params=sampling_params,
+                pooling_params=None,
+                generator=None,
+                block_ids=new_req_data.block_ids,
+                num_computed_tokens=new_req_data.num_computed_tokens,
+                output_token_ids=[],
+                lora_request=new_req_data.lora_request,
+            )
+
+            req_ids_to_add.append(req_id)
+
+            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
+            if self.uses_mrope:
+                image_grid_thw = []
+                video_grid_thw = []
+                second_per_grid_ts = []
+                audio_feature_lengths = []
+                use_audio_in_video = False
+                for mm_feature in self.requests[req_id].mm_features:
+                    item = mm_feature.data
+                    if item is None:
+                        continue
+                    mm_input = item.get_data()
+                    if mm_input.get("image_grid_thw") is not None:
+                        image_grid_thw.append(
+                            mm_input["image_grid_thw"].tolist())
+                    if mm_input.get("video_grid_thw") is not None:
+                        video_grid_thw.append(
+                            mm_input["video_grid_thw"].tolist())
+                    if mm_input.get("second_per_grid_ts") is not None:
+                        second_per_grid_ts.append(
+                            mm_input["second_per_grid_ts"])
+                    if mm_input.get("audio_feature_lengths") is not None:
+                        audio_feature_lengths.append(
+                            mm_input["audio_feature_lengths"])
+                    if mm_input.get("use_audio_in_video") is True:
+                        use_audio_in_video = True
+
+                hf_config = self.model_config.hf_config
+
+                self.requests[req_id].mrope_positions, self.requests[
+                    req_id].mrope_position_delta = get_mrope_input_positions_fn(
+                        self.requests[req_id].prompt_token_ids,
+                        hf_config=hf_config,
+                        image_grid_thw=image_grid_thw,
+                        video_grid_thw=video_grid_thw,
+                        second_per_grid_ts=second_per_grid_ts,
+                        audio_feature_lengths=audio_feature_lengths,
+                        use_audio_in_video=use_audio_in_video,
+                    )
+
+        # Update the states of the running/resumed requests.
+        req_data = scheduler_output.scheduled_cached_reqs
+        for i, req_id in enumerate(req_data.req_ids):
+            req_state = self.requests[req_id]
+            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]
+
+            # Update the cached states.
+            req_state.num_computed_tokens = num_computed_tokens
+            if not resumed_from_preemption:
+                if new_block_ids is not None:
+                    # Append the new blocks to the existing block IDs.
+                    for block_ids, new_ids in zip(req_state.block_ids,
+                                                  new_block_ids):
+                        block_ids.extend(new_ids)
+            else:
+                assert new_block_ids is not None
+                # The request is resumed from preemption.
+                # 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
+                # scheduled in the previous step and needs to be added again.
+                req_ids_to_add.append(req_id)
+                continue
+
+            # Update the persistent batch.
+            self.input_batch.num_computed_tokens_cpu[
+                req_index] = num_computed_tokens
+            if new_block_ids is not None:
+                self.input_batch.block_table.append_row(
+                    new_block_ids, req_index)
+
+            # Add spec_token_ids to token_ids_cpu.
+            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
+                req_id, ())
+            if spec_token_ids:
+                num_spec_tokens = len(spec_token_ids)
+                start_index = self.input_batch.num_tokens_no_spec[req_index]
+                end_token_index = start_index + num_spec_tokens
+                self.input_batch.token_ids_cpu[
+                    req_index, start_index:end_token_index] = spec_token_ids
+                # NOTE(woosuk): `num_tokens` here may include spec tokens.
+                self.input_batch.num_tokens[req_index] += num_spec_tokens
+
+        # Add the new or resumed requests to the persistent batch.
+        # The smaller empty indices are filled first.
+        removed_req_indices = sorted(removed_req_indices, reverse=True)
+        for req_id in req_ids_to_add:
+            req_state = self.requests[req_id]
+            if removed_req_indices:
+                # Fill the empty index.
+                req_index = removed_req_indices.pop()
+            else:
+                # Append to the end.
+                req_index = None
+            self.input_batch.add_request(req_state, req_index)
+
+        # Condense the batched states if there are empty indices.
+        if removed_req_indices:
+            self.input_batch.condense(removed_req_indices)
+
+        batch_changed = len(unscheduled_req_ids) > 0 or len(req_ids_to_add) > 0
+        # TODO(jevinjiang): I assume we do not need to set batch_changed to true if just swapping requests.
+        self._reorder_batch(scheduler_output)
+        return batch_changed