tpu-inference 0.11.1__py3-none-any.whl

tpu_inference/runner/speculative_decoding_manager.py

@@ -0,0 +1,250 @@
+from __future__ import annotations
+
+import functools
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Optional
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+from vllm.v1.core.sched.output import SchedulerOutput as VllmSchedulerOutput
+from vllm.v1.outputs import DraftTokenIds
+from vllm.v1.spec_decode.ngram_proposer import NgramProposer
+
+from tpu_inference.runner import utils as runner_utils
+from tpu_inference.spec_decode.jax.eagle3 import Eagle3Proposer
+from tpu_inference.utils import device_array
+
+if TYPE_CHECKING:
+    from tpu_inference.layers.common.attention_metadata import \
+        AttentionMetadata
+    from tpu_inference.runner.tpu_jax_runner import TPUModelRunner
+
+
+@dataclass
+class SpecDecodeMetadata:
+    """Metadata for speculative decoding on JAX/TPU, containing all necessary indices."""
+    draft_token_ids: jnp.ndarray
+    draft_lengths: jnp.ndarray
+    draft_lengths_cpu: np.ndarray
+    target_logits_indices: jnp.ndarray
+    bonus_logits_indices: jnp.ndarray
+    final_logits_indices: jnp.ndarray
+
+
+class SpeculativeDecodingManager:
+
+    def __init__(self, runner: TPUModelRunner):
+        self.runner = runner
+        # Cached draft tokens.
+        self._draft_token_ids: Optional[list[list[int]]] = None
+
+    def take_draft_token_ids(self) -> Optional[DraftTokenIds]:
+        if self._draft_token_ids is None:
+            return None
+        req_ids = self.runner.input_batch.req_ids
+        draft_token_ids = self._draft_token_ids
+        self._draft_token_ids = None
+        return DraftTokenIds(req_ids, draft_token_ids)
+
+    def propose_draft_token_ids(
+        self,
+        sampled_token_ids: list[list[int]],
+        aux_hidden_states: Optional[tuple[jnp.ndarray, ...]],
+        attn_metadata: AttentionMetadata,
+        spec_decode_metadata: Optional[SpecDecodeMetadata],
+        scheduler_output: Optional[VllmSchedulerOutput] = None,
+        input_ids: Optional[jnp.ndarray] = None,
+    ) -> None:
+        if self.runner.speculative_config.method == "ngram":
+            assert isinstance(self.runner.drafter, NgramProposer)
+            self._draft_token_ids = self.runner.drafter.propose(
+                sampled_token_ids[:self.runner.input_batch.num_reqs],
+                self.runner.input_batch.req_ids,
+                self.runner.input_batch.num_tokens_no_spec,
+                self.runner.input_batch.token_ids_cpu,
+                self.runner.input_batch.spec_decode_unsupported_reqs)
+        elif self.runner.speculative_config.method == "eagle3":
+            self._draft_token_ids = self.propose_eagle3_draft_token_ids(
+                sampled_token_ids,
+                aux_hidden_states,
+                attn_metadata,
+                spec_decode_metadata,
+                scheduler_output,
+                input_ids,
+            )
+        else:
+            raise NotImplementedError(
+                f"Speculative decoding method "
+                f"'{self.runner.speculative_config.method}' is not supported.")
+
+    @functools.partial(jax.jit, static_argnums=(0, ))
+    def _convert_list_to_device_array(self, lst: list[int]) -> jnp.ndarray:
+        """Jitted helper function to convert a list to a device array."""
+        arr = jnp.array(lst, dtype=jnp.int32)
+        return device_array(self.runner.mesh, arr)
+
+    def propose_eagle3_draft_token_ids(
+        self,
+        sampled_token_ids: list[list[int]],
+        aux_hidden_states: Optional[tuple[jnp.ndarray, ...]],
+        attn_metadata: AttentionMetadata,
+        spec_decode_metadata: Optional[SpecDecodeMetadata],
+        scheduler_output: VllmSchedulerOutput,
+        input_ids: jnp.ndarray,
+    ) -> list[list[int]]:
+        assert isinstance(self.runner.drafter, Eagle3Proposer)
+
+        # TODO(woosuk): Refactor the loop.
+        req_ids = self.runner.input_batch.req_ids
+        next_token_ids: list[int] = []
+        for i, token_ids in enumerate(sampled_token_ids):
+            if token_ids:
+                # Common case.
+                next_token_id = token_ids[-1]
+            else:
+                # Partial prefill (rare case).
+                # Get the next token id from the request state.
+                req_id = req_ids[i]
+                req_state = self.runner.requests[req_id]
+                seq_len = (req_state.num_computed_tokens +
+                           scheduler_output.num_scheduled_tokens[req_id])
+                next_token_id = req_state.get_token_id(seq_len)
+            next_token_ids.append(next_token_id)
+
+        # Pad the batch size to match with existing padding for target model
+        pad_len = attn_metadata.seq_lens.shape[0] - len(next_token_ids)
+        assert pad_len >= 0
+        next_token_ids += [0] * pad_len
+
+        next_token_ids = self._convert_list_to_device_array(next_token_ids)
+
+        if spec_decode_metadata is None:
+            num_rejected_tokens = None
+        else:
+            num_draft_tokens = spec_decode_metadata.draft_lengths_cpu
+            num_rejected_tokens = [
+                int(n) + 1 - len(sampled_token_ids[i]) if n > 0 else 0
+                for i, n in enumerate(num_draft_tokens)
+            ]
+
+            pad_len = self.runner.max_num_reqs - len(num_rejected_tokens)
+            num_rejected_tokens += [0] * pad_len
+            num_rejected_tokens = self._convert_list_to_device_array(
+                num_rejected_tokens)
+
+        attn_metadata, target_token_ids, target_hidden_states = self.runner.drafter.prepare_inputs(
+            attn_metadata,
+            input_ids,
+            aux_hidden_states,
+            num_rejected_tokens,
+        )
+        self.runner.kv_caches, draft_token_ids = self.runner.drafter.propose(
+            kv_caches=self.runner.kv_caches,
+            next_token_ids=next_token_ids,
+            attn_metadata=attn_metadata,
+            target_token_ids=target_token_ids,
+            target_hidden_states=target_hidden_states,
+        )
+        result = draft_token_ids.tolist()
+        return result
+
+    def get_spec_decode_metadata(
+        self,
+        num_draft_tokens: np.ndarray,
+        cu_num_scheduled_tokens: np.ndarray,
+        padded_num_reqs: int,
+    ) -> SpecDecodeMetadata:
+        # Inputs:
+        # cu_num_scheduled_tokens:  [  4, 104, 107, 207, 209]
+        # num_draft_tokens:         [  3,   0,   2,   0,   1]
+        # Outputs:
+        # cu_num_draft_tokens:      [  3,   3,   5,   5,   6]
+        # logits_indices:           [  0,   1,   2,   3, 103, 104, 105, 106,
+        #                            206, 207, 208]
+        # target_logits_indices:    [  0,   1,   2,   5,   6,   9]
+        # bonus_logits_indices:     [  3,   4,   7,   8,  10]
+
+        # Compute the logits indices.
+        # [4, 1, 3, 1, 2]
+        num_sampled_tokens = num_draft_tokens + 1
+
+        # Step 1. cu_num_sampled_tokens: [4, 5, 8, 9, 11]
+        # arange: [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
+        cu_num_sampled_tokens = np.cumsum(num_sampled_tokens)
+        arange = np.concatenate(
+            [self.runner.arange_cpu[:n] for n in num_sampled_tokens])
+        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
+        logits_indices = np.repeat(
+            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens)
+        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
+        logits_indices += arange
+        # Compute the bonus logits indices.
+        bonus_logits_indices = cu_num_sampled_tokens - 1
+
+        # Compute the draft logits indices.
+        # arange: [0, 1, 2, 0, 1, 0]
+        arange = np.concatenate(
+            [self.runner.arange_cpu[:n] for n in num_draft_tokens])
+        # [0, 0, 0, 5, 5, 9]
+        target_logits_indices = np.repeat(
+            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens)
+        # [0, 1, 2, 5, 6, 9]
+        target_logits_indices += arange
+
+        # Compute the draft token ids.
+        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
+        draft_token_ids = self.runner.input_ids_cpu[logits_indices]
+        draft_token_ids = draft_token_ids[target_logits_indices + 1]
+        padded_logits_length = runner_utils.get_padded_token_len(
+            self.runner.num_logits_paddings, logits_indices.shape[0])
+        padded_logits_indices = np.concatenate([
+            logits_indices,
+            np.zeros(padded_logits_length - logits_indices.shape[0],
+                     dtype=np.int32)
+        ])
+
+        assert bonus_logits_indices.shape[0] <= padded_num_reqs, (
+            f"bonus_logits_indices.shape[0]={bonus_logits_indices.shape[0]} "
+            f"padded_num_reqs={padded_num_reqs}")
+
+        padded_bonus_logits_indices = np.concatenate([
+            bonus_logits_indices,
+            np.zeros(padded_num_reqs - bonus_logits_indices.shape[0],
+                     dtype=np.int32)
+        ])
+        padded_num_draft_tokens = np.concatenate([
+            num_draft_tokens,
+            np.zeros(padded_num_reqs - num_draft_tokens.shape[0],
+                     dtype=np.int32)
+        ])
+        padded_draft_token_ids = np.concatenate([
+            draft_token_ids,
+            np.zeros(padded_logits_length - draft_token_ids.shape[0],
+                     dtype=np.int32)
+        ])
+        padded_target_logits_indices = np.concatenate([
+            target_logits_indices,
+            np.zeros(padded_logits_length - target_logits_indices.shape[0],
+                     dtype=np.int32)
+        ])
+
+        padded_num_draft_tokens_cpu = padded_num_draft_tokens
+        # CPU -> TPU copy.
+        (padded_num_draft_tokens, padded_draft_token_ids,
+         padded_logits_indices, padded_target_logits_indices,
+         padded_bonus_logits_indices) = device_array(
+             self.runner.mesh,
+             (padded_num_draft_tokens, padded_draft_token_ids,
+              padded_logits_indices, padded_target_logits_indices,
+              padded_bonus_logits_indices))
+
+        metadata = SpecDecodeMetadata(
+            draft_token_ids=padded_draft_token_ids,
+            draft_lengths=padded_num_draft_tokens,
+            draft_lengths_cpu=padded_num_draft_tokens_cpu,
+            target_logits_indices=padded_target_logits_indices,
+            bonus_logits_indices=padded_bonus_logits_indices,
+            final_logits_indices=padded_logits_indices,
+        )
+        return metadata

tpu_inference/runner/structured_decoding_manager.py

@@ -0,0 +1,89 @@
+import functools
+from typing import TYPE_CHECKING, Tuple
+
+import jax
+import jax.numpy as jnp
+
+from tpu_inference.utils import device_array
+
+if TYPE_CHECKING:
+    from vllm.v1.core.sched.output import \
+        SchedulerOutput as VllmSchedulerOutput
+
+    from tpu_inference.runner.tpu_jax_runner import TPUModelRunner
+
+
+class StructuredDecodingManager:
+
+    def __init__(self, runner: "TPUModelRunner"):
+        self.runner = runner
+
+    @functools.partial(jax.jit, static_argnums=(0, ))
+    def structured_decode_fn(self, require_struct_decoding: jax.Array,
+                             grammar_bitmask: jax.Array, logits: jax.Array,
+                             arange: jax.Array) -> jax.Array:
+        return jax.lax.cond(
+            jnp.any(require_struct_decoding),
+            lambda: self._apply_grammar_bitmask_kernel(
+                logits, grammar_bitmask, require_struct_decoding, arange),
+            lambda: logits)
+
+    @functools.partial(jax.jit, static_argnums=(0, ))
+    def _apply_grammar_bitmask_kernel(self, logits: jax.Array,
+                                      grammar_bitmask: jax.Array,
+                                      require_struct_decoding: jax.Array,
+                                      arange: jax.Array) -> jax.Array:
+
+        # Unpack the bitmask for the entire batch at once.
+        # grammar_bitmask: (B, N) where B=num_reqs, N=cdiv(vocab_size, 32)
+        # arange: (32,)
+        # (B, N, 1) and (1, 1, 32) broadcast to (B, N, 32)
+        unpacked_bitmask = jnp.right_shift(grammar_bitmask[:, :, None],
+                                           arange[None, None, :]) & 1 == 0
+
+        # Reshape to (B, vocab_size) and apply to logits.
+        # (B, N * 32) -> (B, vocab_size)
+        unpacked_bitmask = unpacked_bitmask.reshape(
+            logits.shape[0], -1)[:, :self.runner.vocab_size]
+
+        masked_logits = jnp.where(unpacked_bitmask, -jnp.inf, logits)
+
+        return jnp.where(require_struct_decoding, masked_logits, logits)
+
+    def prepare_structured_decoding_input(
+        self, logits: jax.Array, scheduler_output: "VllmSchedulerOutput"
+    ) -> Tuple[jax.Array, jax.Array, jax.Array]:
+        grammar_bitmask = scheduler_output.grammar_bitmask
+        assert grammar_bitmask is not None
+        num_reqs, _ = logits.shape
+
+        # Reset pre-allocated tensors
+        self.runner.grammar_bitmask_cpu.fill(0)
+        self.runner.require_structured_out_cpu.fill(0)
+
+        sorted_struct_requests = sorted(
+            scheduler_output.structured_output_request_ids.items(),
+            key=lambda item: item[1])
+
+        cumulative_mask_idx = 0
+        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]
+            self.runner.grammar_bitmask_cpu[batch_index] = grammar_bitmask[
+                cumulative_mask_idx]
+            # It's not guaranteed that all requests in this batch require
+            # structured output, so create a bool tensor to represent
+            # the requests that need structured output.
+            self.runner.require_structured_out_cpu[batch_index] = True
+            cumulative_mask_idx += 1
+
+        (require_structured_out_cpu,
+         grammar_bitmask_cpu, structured_decode_arange) = device_array(
+             self.runner.mesh,
+             (self.runner.require_structured_out_cpu[:num_reqs],
+              self.runner.grammar_bitmask_cpu[:num_reqs],
+              self.runner.structured_decode_arange))
+
+        return (require_structured_out_cpu, grammar_bitmask_cpu,
+                structured_decode_arange)