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

tpu_inference/layers/jax/sample/rejection_sampler.py

@@ -0,0 +1,515 @@
+"""
+JAX-based rejection sampler for speculative decoding on TPU.
+
+This implementation follows the same algorithm as the GPU version but is
+designed for JAX/TPU compatibility. It currently only supports greedy sampling.
+"""
+
+import functools
+from typing import Optional
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+
+from tpu_inference.layers.common.binary_search import topk_mask, topp_mask
+from tpu_inference.layers.jax.sample.sampling_metadata import \
+    TPUSupportedSamplingMetadata
+
+# Placeholder token ID for rejected tokens
+PLACEHOLDER_TOKEN_ID = -1
+GREEDY_TEMPERATURE = -1
+
+
+class RejectionSampler:
+    """
+    JAX-based rejection sampler for speculative decoding.
+
+    The implementation follows the algorithm described in
+    https://arxiv.org/abs/2211.17192.
+    """
+
+    def __init__(self):
+        pass
+
+    def __call__(
+        self,
+        # [num_tokens] - flattened format
+        draft_token_ids: jnp.ndarray,
+        # [batch_size] - number of draft tokens per request
+        num_draft_tokens: jnp.ndarray,
+        # [num_tokens, vocab_size] - flattened format
+        draft_probs: Optional[jnp.ndarray],
+        # [num_tokens, vocab_size] - flattened format
+        target_logits: jnp.ndarray,
+        # [batch_size]
+        bonus_token_ids: jnp.ndarray,
+        sampling_metadata: TPUSupportedSamplingMetadata,
+        key: Optional[jax.random.PRNGKey] = None,
+    ) -> jnp.ndarray:
+        """
+        Perform rejection sampling on draft tokens with flattened inputs.
+
+        Args:
+            draft_token_ids: Draft token IDs in flattened format [num_tokens].
+            num_draft_tokens: Number of draft tokens per request [batch_size].
+            draft_probs: Draft probabilities in flattened format [num_tokens, vocab_size].
+            target_probs: Target probabilities in flattened format [num_tokens, vocab_size].
+            bonus_token_ids: Bonus token IDs [batch_size].
+            sampling_metadata: Additional metadata needed for sampling.
+            key: JAX random key for non-greedy sampling.
+
+        Returns:
+            output_token_ids: A tensor containing the final output token IDs.
+        """
+        return self.forward(
+            draft_token_ids=draft_token_ids,
+            num_draft_tokens=num_draft_tokens,
+            draft_probs=draft_probs,
+            target_logits=target_logits,
+            bonus_token_ids=bonus_token_ids,
+            sampling_metadata=sampling_metadata,
+            key=key,
+        )
+
+    @functools.partial(jax.jit, static_argnums=(0, ))
+    def forward(
+        self,
+        # [num_tokens] - flattened format
+        draft_token_ids: jnp.ndarray,
+        # [batch_size] - number of draft tokens per request
+        num_draft_tokens: jnp.ndarray,
+        # [num_tokens, vocab_size] - flattened format
+        draft_probs: Optional[jnp.ndarray],
+        # [num_tokens, vocab_size] - flattened format
+        target_logits: jnp.ndarray,
+        # [batch_size]
+        bonus_token_ids: jnp.ndarray,
+        sampling_metadata: TPUSupportedSamplingMetadata,
+        key: Optional[jax.random.PRNGKey] = None,
+    ) -> jnp.ndarray:
+        """
+        Perform rejection sampling on draft tokens with flattened inputs.
+
+        Args:
+            draft_token_ids: Draft token IDs in flattened format [num_tokens].
+            num_draft_tokens: Number of draft tokens per request [batch_size].
+            draft_probs: Draft probabilities in flattened format [num_tokens, vocab_size].
+            target_logits: Target logits in flattened format [num_tokens, vocab_size].
+            bonus_token_ids: Bonus token IDs [batch_size].
+            sampling_metadata: Additional metadata needed for sampling.
+            key: JAX random key for non-greedy sampling.
+
+        Returns:
+            output_token_ids: A tensor containing the final output token IDs.
+        """
+
+        if sampling_metadata.do_sampling:
+            target_probs = _compute_probs(target_logits, num_draft_tokens,
+                                          sampling_metadata)
+        else:
+            target_probs = target_logits
+
+        output_token_ids = rejection_sample(
+            draft_token_ids,
+            num_draft_tokens,
+            draft_probs,
+            target_probs,
+            bonus_token_ids,
+            sampling_metadata,
+            key=key,
+        )
+        return output_token_ids
+
+    @staticmethod
+    def parse_output(
+        output_token_ids: jnp.ndarray,
+        vocab_size: int,
+        num_draft_tokens_cpu: np.ndarray,
+        batch_size: int,
+        padded_tokens_length: int,
+    ) -> list[list[int]]:
+        """Parse the output of the rejection sampler.
+
+        Args:
+            output_token_ids: The sampled token IDs in shape
+                [num_tokens + batch_size]. The first num_tokens elements are
+                the main tokens, and the last batch_size elements are bonus tokens.
+                Rejected tokens are replaced with `PLACEHOLDER_TOKEN_ID`.
+            vocab_size: The size of the vocabulary.
+            num_draft_tokens_cpu: Number of draft tokens per request [batch_size]
+                as a numpy array on CPU.
+            batch_size: The number of requests in the batch.
+            padded_tokens_length: The padded length of the main tokens in the output.
+
+        Returns:
+            A list of lists of token IDs.
+        """
+        # Convert JAX array to numpy for easier manipulation
+        output_token_ids_np = np.asarray(output_token_ids)
+
+        # Split main tokens and bonus tokens
+        main_tokens = output_token_ids_np[:
+                                          padded_tokens_length]  # [num_tokens]
+        bonus_tokens = output_token_ids_np[
+            padded_tokens_length:]  # [batch_size]
+
+        # Reconstruct per-sequence outputs
+        outputs = []
+        start_idx = 0
+
+        for i in range(batch_size):
+            seq_length = int(num_draft_tokens_cpu[i])
+            end_idx = start_idx + seq_length
+
+            # Get main tokens for this sequence
+            seq_main_tokens = main_tokens[start_idx:end_idx]
+
+            # Filter out placeholder tokens
+            valid_main_tokens = seq_main_tokens[
+                (seq_main_tokens != PLACEHOLDER_TOKEN_ID)
+                & (seq_main_tokens < vocab_size)]
+
+            # Add bonus token if it's valid
+            bonus_token = bonus_tokens[i]
+            if bonus_token != PLACEHOLDER_TOKEN_ID and bonus_token < vocab_size:
+                seq_tokens = np.concatenate([valid_main_tokens, [bonus_token]])
+            else:
+                seq_tokens = valid_main_tokens
+
+            outputs.append(seq_tokens.tolist())
+            start_idx = end_idx
+
+        return outputs
+
+
+def _compute_probs(
+    logits: jnp.ndarray,
+    num_draft_tokens: jnp.ndarray,
+    sampling_metadata: TPUSupportedSamplingMetadata,
+) -> jnp.ndarray:
+    """
+    Apply top-k, top-p, and temperature to logits and compute probabilities.
+    """
+    total_tokens = logits.shape[0]
+    segment_ids, _ = _get_segment_info(num_draft_tokens, total_tokens)
+
+    # Expand sampling params from [batch_size] to [num_tokens]
+    top_k = sampling_metadata.top_k[segment_ids]
+    top_p = sampling_metadata.top_p[segment_ids]
+    temperatures = sampling_metadata.temperature[segment_ids]
+
+    # Apply top-k and top-p masking
+    logits = logits.astype(jnp.float32)
+    # Only apply top-k masking if k > 0 for each token
+    should_apply_topk = jnp.expand_dims(top_k > 0, axis=-1)
+    topk_masked = topk_mask(logits, top_k, replace_val=-jnp.inf)
+    logits = jnp.where(should_apply_topk, topk_masked, logits)
+
+    # Only apply top-p masking if p < 1.0 for each token
+    should_apply_topp = jnp.expand_dims(top_p < 1.0, axis=-1)
+    topp_masked = topp_mask(logits, top_p, replace_val=-jnp.inf)
+    logits = jnp.where(should_apply_topp, topp_masked, logits)
+
+    # Apply temperature scaling
+    temperatures = jnp.expand_dims(temperatures, axis=-1)
+    # Add epsilon to avoid division by zero
+    logits /= (temperatures + 1e-9)
+
+    return jax.nn.softmax(logits, axis=-1)
+
+
+def _get_segment_info(num_draft_tokens: jax.Array, total_tokens: int):
+    """Helper to create segment IDs and per-segment indices."""
+    batch_size = num_draft_tokens.shape[0]
+
+    # `segment_ids` assigns a unique ID to each token, corresponding to its
+    # sequence in the batch. E.g., [0, 0, 0, 1, 1, 2, 2, 2, 2] for sequences [3, 2, 4].
+    segment_ids = jnp.repeat(jnp.arange(batch_size),
+                             num_draft_tokens,
+                             total_repeat_length=total_tokens)
+
+    # `group_indices` creates a within-segment index for each token.
+    # E.g., [0, 1, 2, 0, 1, 0, 1, 2, 3] for the example above.
+    segment_starts = jnp.concatenate(
+        [jnp.array([0]), jnp.cumsum(num_draft_tokens)[:-1]])
+    broadcast_starts = jnp.repeat(segment_starts,
+                                  num_draft_tokens,
+                                  total_repeat_length=total_tokens)
+    group_indices = jnp.arange(total_tokens) - broadcast_starts
+    return segment_ids, group_indices
+
+
+def _sample_recovered_tokens(
+    draft_token_ids: jax.Array,
+    draft_probs: Optional[jax.Array],
+    target_probs: jax.Array,
+    key: jax.random.PRNGKey,
+) -> jax.Array:
+    """
+    Sample recovered tokens using the Gumbel-Max trick.
+    This is used when a draft token is rejected in random sampling.
+    """
+    if draft_probs is not None:
+        # The new distribution is p' = max(p_target - p_draft, 0)
+        new_dist = jnp.maximum(target_probs - draft_probs, 0)
+    else:
+        # If no draft probs, the new distribution is the target distribution
+        # with the draft token's probability zeroed out.
+        vocab_size = target_probs.shape[-1]
+        mask = jax.nn.one_hot(draft_token_ids, vocab_size, dtype=jnp.bool)
+        new_dist = target_probs * ~mask
+
+    # Gumbel-Max trick to sample from the new distribution:
+    # y = argmax(log(p') + g) where g ~ Gumbel(0,1)
+    # This is equivalent to argmax(p' / q) where q ~ Exponential(1)
+    q = jax.random.exponential(key, shape=new_dist.shape)
+
+    # Add a small epsilon to avoid division by zero
+    recovered_token_ids = jnp.argmax(new_dist / (q + 1e-9), axis=-1)
+    return recovered_token_ids
+
+
+def rejection_sample(
+    # [num_tokens] - flattened format
+    draft_token_ids: jnp.ndarray,
+    # [batch_size] - JAX array
+    num_draft_tokens: jnp.ndarray,
+    # [num_tokens, vocab_size] - flattened format
+    draft_probs: Optional[jnp.ndarray],
+    # [num_tokens, vocab_size] - flattened format
+    target_probs: jnp.ndarray,
+    # [batch_size]
+    bonus_token_ids: jnp.ndarray,
+    sampling_metadata: TPUSupportedSamplingMetadata,
+    key: Optional[jax.random.PRNGKey] = None,
+) -> jnp.ndarray:
+    """
+    Perform rejection sampling on draft tokens with flattened inputs.
+
+    Args:
+        draft_token_ids: Draft token IDs in flattened format [num_tokens].
+        num_draft_tokens: Number of draft tokens per request [batch_size].
+        draft_probs: Draft probabilities in flattened format [num_tokens, vocab_size].
+        target_probs: Target probabilities in flattened format [num_tokens, vocab_size].
+        bonus_token_ids: Bonus token IDs [batch_size].
+        sampling_metadata: Sampling metadata.
+        key: JAX random key for non-greedy sampling.
+
+    Returns:
+        output_token_ids: Output token IDs [num_tokens + batch_size].
+    """
+    if sampling_metadata.do_sampling is False:
+        greedy_output = _greedy_rejection_sample_with_segment(
+            draft_token_ids, target_probs, num_draft_tokens, bonus_token_ids)
+        return greedy_output
+
+    # Random path
+    if key is None:
+        raise ValueError(
+            "A random key must be provided for non-greedy sampling.")
+
+    random_output = _random_rejection_sample_with_segment(
+        draft_token_ids,
+        draft_probs,
+        target_probs,
+        num_draft_tokens,
+        bonus_token_ids,
+        key,
+    )
+
+    return random_output
+
+
+def _random_rejection_sample_with_segment(
+    draft_token_ids: jax.Array,
+    draft_probs: Optional[jax.Array],
+    target_probs: jax.Array,
+    num_draft_tokens: jax.Array,
+    bonus_token_ids: jax.Array,
+    key: jax.random.PRNGKey,
+) -> jax.Array:
+    """
+    Performs random speculative decoding validation in a vectorized, jittable manner.
+    """
+    total_tokens = draft_token_ids.shape[0]
+    batch_size = num_draft_tokens.shape[0]
+
+    # Split random key
+    uniform_key, recover_key = jax.random.split(key)
+
+    # --- Step 1: Get Segment Info ---
+    segment_ids, group_indices = _get_segment_info(num_draft_tokens,
+                                                   total_tokens)
+
+    # --- Step 2: Acceptance/Rejection Logic ---
+    if draft_probs is not None:
+        draft_token_probs = jnp.take_along_axis(draft_probs,
+                                                draft_token_ids[:, None],
+                                                axis=-1).squeeze(-1)
+    else:
+        draft_token_probs = 1.0
+
+    target_token_probs = jnp.take_along_axis(target_probs,
+                                             draft_token_ids[:, None],
+                                             axis=-1).squeeze(-1)
+
+    uniform_probs = jax.random.uniform(uniform_key, shape=(total_tokens, ))
+
+    # Acceptance condition: p_target(d) / p_draft(d) >= u
+    ratio = target_token_probs / (draft_token_probs + 1e-9)
+    accepted = ratio >= uniform_probs
+
+    # --- Step 3: Find First Rejection ---
+    rejections = ~accepted
+    large_value = total_tokens
+    rejection_indices = jnp.where(rejections, group_indices, large_value)
+
+    first_rejection_idx_per_segment = jax.ops.segment_min(
+        data=rejection_indices.astype(jnp.int32),
+        segment_ids=segment_ids,
+        num_segments=batch_size,
+        indices_are_sorted=True,
+    )
+
+    max_int = jnp.iinfo(jnp.int32).max
+    first_rejection_idx_per_segment = jnp.where(
+        first_rejection_idx_per_segment == max_int, large_value,
+        first_rejection_idx_per_segment)
+
+    # --- Step 4: Sample Recovered Tokens ---
+    recovered_token_ids = _sample_recovered_tokens(draft_token_ids,
+                                                   draft_probs, target_probs,
+                                                   recover_key)
+
+    # --- Step 5: Generate Main Token Output ---
+    first_rejection_idx_broadcast = jnp.repeat(
+        first_rejection_idx_per_segment,
+        num_draft_tokens,
+        total_repeat_length=total_tokens)
+
+    main_tokens = jnp.where(
+        group_indices < first_rejection_idx_broadcast, draft_token_ids,
+        jnp.where(group_indices == first_rejection_idx_broadcast,
+                  recovered_token_ids, PLACEHOLDER_TOKEN_ID))
+
+    # --- Step 6: Handle Bonus Tokens ---
+    all_accepted = first_rejection_idx_per_segment == large_value
+    no_draft_tokens = num_draft_tokens == 0
+    should_get_bonus = all_accepted | no_draft_tokens
+    bonus_tokens = jnp.where(should_get_bonus, bonus_token_ids,
+                             PLACEHOLDER_TOKEN_ID)
+
+    # --- Step 7: Concatenate ---
+    return jnp.concatenate([main_tokens, bonus_tokens])
+
+
+# TODO(pooyam): Optimize/Profile this implementation further. Currently, I just want working e2e. There might be overheads with `parse_output` that can be optimized on TPU.
+# I should Benchmark against the following approaches:
+# - Using `jax.lax.segment_xyz`` to work with flattened inputs instead of batched inputs.
+# - Using vectorized implementation using `cumprod` and other masking tricks.
+# - A pallas kernel similar to the Triton implementation.
+# - Scan based approach.
+# Overall, I expect XLA to optimize the scan-based approach pretty well, but
+# it would be good to compare performance against other methods.
+def _greedy_rejection_sample_with_segment(
+    draft_token_ids: jax.Array,
+    target_probs: jax.Array,
+    num_draft_tokens: jax.Array,
+    bonus_token_ids: jax.Array,
+) -> jax.Array:
+    """
+    Performs greedy speculative decoding validation in a vectorized, jittable manner.
+
+    This function compares draft tokens with the target model's outputs. For each
+    sequence in the batch, it accepts tokens as long as the draft and target match.
+    When a mismatch occurs, it takes the target model's token and invalidates the
+    rest of the tokens in that sequence by setting them to -1.
+
+    Args:
+        draft_token_ids: A 1D JAX array (num_tokens,) of integers representing the
+                         concatenated draft tokens for all sequences in the batch.
+        target_probs: A 2D JAX array (num_tokens, vocab_size) of floats representing
+                      the concatenated target model's probabilities.
+        num_draft_tokens: A 1D JAX array (batch_size,) of integers specifying the
+                          number of draft tokens for each sequence in the batch.
+        bonus_token_ids: A 1D JAX array (batch_size,) of integers representing the
+                         bonus token for each sequence.
+
+    Returns:
+        A 1D JAX array (num_tokens + batch_size,) containing the validated token
+        sequence followed by bonus tokens (or -1 if not accepted).
+    """
+    # Get target argmax
+    target_logits_argmax = jnp.argmax(target_probs, axis=-1)
+
+    # --- Step 1: Create Segment IDs and Per-Segment Indices ---
+    total_tokens = draft_token_ids.shape[0]
+    batch_size = num_draft_tokens.shape[0]
+    segment_ids, group_indices = _get_segment_info(num_draft_tokens,
+                                                   total_tokens)
+
+    # --- Step 2: Find the First Mismatch in Each Segment ---
+
+    # Find all mismatches between draft and target tokens.
+    mismatches = draft_token_ids != target_logits_argmax
+
+    # To find the *first* mismatch, we use a trick with segment_min.
+    # We create an array where mismatched positions hold their `group_index`
+    # and matched positions hold a large value.
+    large_value = total_tokens
+    mismatch_indices = jnp.where(mismatches, group_indices, large_value)
+
+    # `segment_min` finds the minimum `mismatch_index` for each segment. This
+    # effectively gives us the `group_index` of the first mismatch.
+    # For sequences with no mismatches, the result will be `large_value`.
+    first_mismatch_idx_per_segment = jax.ops.segment_min(
+        data=mismatch_indices.astype(jnp.int32),
+        segment_ids=segment_ids,
+        num_segments=batch_size,
+        indices_are_sorted=True,
+    )
+
+    # Handle empty segments (where num_draft_tokens is 0). `segment_min` returns
+    # the dtype's max value for empty segments; we replace it with our large_value
+    # for consistency.
+    max_int = jnp.iinfo(jnp.int32).max
+    first_mismatch_idx_per_segment = jnp.where(
+        first_mismatch_idx_per_segment == max_int, large_value,
+        first_mismatch_idx_per_segment)
+
+    # --- Step 3: Broadcast Mismatch Info and Generate Main Token Output ---
+
+    # Broadcast the first mismatch index back to the original token dimension.
+    first_mismatch_idx_broadcast = jnp.repeat(first_mismatch_idx_per_segment,
+                                              num_draft_tokens,
+                                              total_repeat_length=total_tokens)
+
+    # The final logic for main tokens:
+    # A token is valid if its `group_index` is less than or equal to the
+    # index of the first mismatch in its segment.
+    # - If `group_index < first_mismatch_idx`, the draft was correct.
+    # - If `group_index == first_mismatch_idx`, this is the correction token.
+    # - If `group_index > first_mismatch_idx`, the token is invalid (-1).
+    main_tokens = jnp.where(group_indices <= first_mismatch_idx_broadcast,
+                            target_logits_argmax, PLACEHOLDER_TOKEN_ID)
+
+    # --- Step 4: Handle Bonus Tokens ---
+
+    # A sequence gets its bonus token if there were no mismatches
+    # (first_mismatch_idx_per_segment == large_value)
+    all_accepted = first_mismatch_idx_per_segment == large_value
+
+    # For sequences with no draft tokens, we should still give them the bonus token
+    # since there's nothing to reject
+    no_draft_tokens = num_draft_tokens == 0
+    should_get_bonus = all_accepted | no_draft_tokens
+
+    bonus_tokens = jnp.where(should_get_bonus, bonus_token_ids,
+                             PLACEHOLDER_TOKEN_ID)
+
+    # --- Step 5: Concatenate Main Tokens and Bonus Tokens ---
+
+    output = jnp.concatenate([main_tokens, bonus_tokens])
+
+    return output

tpu_inference/layers/jax/sample/sampling.py

@@ -0,0 +1,96 @@
+import functools
+
+import jax
+import jax.numpy as jnp
+from jax.sharding import Mesh, NamedSharding
+from jax.sharding import PartitionSpec as P
+from vllm.v1.outputs import LogprobsTensors
+
+from tpu_inference.layers.common.binary_search import topk_mask, topp_mask
+from tpu_inference.layers.common.sharding import ShardingAxisName
+from tpu_inference.layers.jax.sample.sampling_metadata import \
+    TPUSupportedSamplingMetadata
+
+_SAMPLING_EPS = 1e-5
+
+
+@functools.partial(
+    jax.jit,
+    static_argnames=["mesh"],
+)
+def sample(
+    rng: jax.Array,
+    mesh: Mesh,
+    logits: jax.Array,
+    tpu_sampling_metadata: TPUSupportedSamplingMetadata,
+) -> jax.Array:
+    # (B, vocab_size)
+    if tpu_sampling_metadata.do_sampling:
+        # Unshard the logits explicity to avoid latency increase.
+        logits = jax.lax.with_sharding_constraint(
+            logits, NamedSharding(mesh, P(ShardingAxisName.ATTN_DATA, None)))
+    greedy_sampled = jnp.argmax(logits, axis=-1)
+    if not tpu_sampling_metadata.do_sampling:
+        return greedy_sampled
+
+    logits = logits.astype(jnp.float32)
+    logits = topk_mask(logits, tpu_sampling_metadata.top_k, replace_val=-1e12)
+    logits = topp_mask(logits, tpu_sampling_metadata.top_p, replace_val=-1e12)
+
+    temperatures = tpu_sampling_metadata.temperature.astype(logits.dtype)
+    temperatures = jnp.expand_dims(temperatures, axis=-1)
+    logits /= temperatures
+
+    # (batch_size,)
+    next_tokens = jax.random.categorical(rng, logits)
+    # Note: avoid using the sample result when temperature < _SAMPLING_EPS
+    # If temperature < 0, logits /= temperatures will flip the result, causing error.
+    return jnp.where(tpu_sampling_metadata.temperature < _SAMPLING_EPS,
+                     greedy_sampled, next_tokens)
+
+
+def compute_logprobs(logits: jax.Array) -> jax.Array:
+    return jax.nn.log_softmax(logits, axis=-1)
+
+
+def gather_logprobs(
+    logprobs: jax.Array,
+    token_ids: jax.Array,
+    num_logprobs: int,
+) -> LogprobsTensors:
+    """
+    Gather logprobs for topk and sampled/prompt token.
+
+    Args:
+        logprobs: (num tokens) x (vocab) tensor
+        token_ids: prompt tokens (if prompt logprobs)
+                    or sampled tokens (if sampled
+                    logprobs); 1D token ID tensor
+                    with (num tokens) elements
+        num_logprobs: minimum number of logprobs to
+                    retain per token
+
+
+    Returns:
+        Top-k int indices tensor, (num tokens) x (num_logprobs + 1)
+        Top-k float logprobs tensor, (num tokens) x (num_logprobs + 1)
+        Sampled token rank tensor, (num tokens)
+    """
+    # Find the topK values.
+    topk_logprobs, topk_indices = jax.lax.top_k(logprobs, k=num_logprobs)
+
+    # Get with the logprob of the prompt or sampled token.
+    token_ids = jnp.expand_dims(token_ids, axis=-1)
+    token_logprobs = jnp.take_along_axis(logprobs, token_ids, axis=-1)
+
+    # Compute the ranks of the actual token.
+    token_ranks = jnp.sum(logprobs >= token_logprobs, axis=-1)
+
+    # Concatenate together with the topk.
+    indices = jnp.concatenate((token_ids, topk_indices), axis=1)
+    logprobs = jnp.concatenate((token_logprobs, topk_logprobs), axis=1)
+
+    # Use int32 to reduce the tensor size.
+    indices = jnp.int32(indices)
+
+    return LogprobsTensors(indices, logprobs, token_ranks)

tpu_inference/layers/jax/sample/sampling_metadata.py

@@ -0,0 +1,76 @@
+import functools
+from dataclasses import dataclass
+from typing import Optional
+
+import jax
+import jax.numpy as jnp
+import torch
+from jax.sharding import Mesh
+
+from tpu_inference.runner.input_batch import InputBatch
+from tpu_inference.utils import device_array
+
+DEFAULT_SAMPLING_PARAMS = dict(
+    temperature=-1.0,
+    top_k=0,
+    top_p=1.0,
+)
+
+
+@functools.partial(
+    jax.tree_util.register_dataclass,
+    data_fields=[
+        "temperature",
+        "top_k",
+        "top_p",
+    ],
+    meta_fields=["do_sampling", "logprobs"],
+)
+@dataclass
+class TPUSupportedSamplingMetadata:
+    temperature: Optional[jnp.ndarray] = None
+    top_k: Optional[jnp.ndarray] = None
+    top_p: Optional[jnp.ndarray] = None
+    do_sampling: bool = False
+    logprobs: bool = False
+
+    @classmethod
+    def from_input_batch(
+        cls,
+        mesh: Mesh,
+        input_batch: InputBatch,
+        padded_num_reqs: int,
+        sharding: Optional[jax.sharding.Sharding] = None,
+    ) -> "TPUSupportedSamplingMetadata":
+        needs_logprobs = input_batch.max_num_logprobs > 0 if input_batch.max_num_logprobs else False
+        if input_batch.all_greedy:
+            return cls(do_sampling=False, logprobs=needs_logprobs)
+        num_reqs = input_batch.num_reqs
+
+        def fill_slice(cpu_torch_tensor: torch.Tensor,
+                       fill_val: float) -> torch.Tensor:
+            # Pad value is the default one.
+            cpu_torch_tensor[num_reqs:padded_num_reqs] = fill_val
+            return cpu_torch_tensor
+
+        temp_tensor = fill_slice(input_batch.temperature_cpu,
+                                 DEFAULT_SAMPLING_PARAMS["temperature"])
+        top_k_tensor = fill_slice(input_batch.top_k_cpu,
+                                  DEFAULT_SAMPLING_PARAMS["top_k"])
+        top_p_tensor = fill_slice(input_batch.top_p_cpu,
+                                  DEFAULT_SAMPLING_PARAMS["top_p"])
+
+        # Slice persistent device tensors to a fixed pre-compiled padded shape.
+        return cls(
+            temperature=device_array(mesh,
+                                     temp_tensor[:padded_num_reqs],
+                                     sharding=sharding),
+            top_p=device_array(mesh,
+                               top_p_tensor[:padded_num_reqs],
+                               sharding=sharding),
+            top_k=device_array(mesh,
+                               top_k_tensor[:padded_num_reqs],
+                               sharding=sharding),
+            do_sampling=not input_batch.all_greedy,
+            logprobs=needs_logprobs,
+        )