[Model Runner V2] Support penalties using bin counts (#29703)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

vllm/v1/worker/gpu/input_batch.py

@@ -341,6 +341,8 @@ def _post_update_kernel(
     idx_mapping_ptr,
     num_computed_tokens_ptr,
     last_sampled_tokens_ptr,
+    output_bin_counts_ptr,
+    output_bin_counts_stride,
     sampled_tokens_ptr,
     sampled_tokens_stride,
     num_sampled_ptr,
@@ -357,6 +359,15 @@ def _post_update_kernel(
         )
         tl.store(last_sampled_tokens_ptr + req_state_idx, token_id)
 
+    for i in range(num_sampled):
+        token_id = tl.load(sampled_tokens_ptr + req_id * sampled_tokens_stride + i)
+        token_ptr = (
+            output_bin_counts_ptr + req_state_idx * output_bin_counts_stride + token_id
+        )
+        count = tl.load(token_ptr)
+        count += 1
+        tl.store(token_ptr, count)
+
     query_start = tl.load(query_start_loc_ptr + req_id)
     query_end = tl.load(query_start_loc_ptr + req_id + 1)
     query_len = query_end - query_start
@@ -374,6 +385,8 @@ def post_update(
     num_computed_tokens: torch.Tensor,
     # [max_num_reqs]
     last_sampled_tokens: torch.Tensor,
+    # [max_num_reqs, vocab_size]
+    output_bin_counts: torch.Tensor,
     # [num_reqs, num_speculative_steps + 1]
     sampled_tokens: torch.Tensor,
     # [num_reqs]
@@ -388,6 +401,8 @@ def post_update(
         idx_mapping,
         num_computed_tokens,
         last_sampled_tokens,
+        output_bin_counts,
+        output_bin_counts.stride(0),
         sampled_tokens,
         sampled_tokens.stride(0),
         num_sampled,

vllm/v1/worker/gpu/model_runner.py

@@ -512,7 +512,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
             idx_mapping_np,
             num_scheduled_tokens,
             query_start_loc_np,
-            self.req_states.prefill_token_ids,
+            self.req_states.prefill_token_ids.np,
             self.req_states.num_computed_prefill_tokens,
             self.input_buffers.input_ids.np,
         )
@@ -681,7 +681,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         # Handle chunked prompts.
         pos_after_step = computed_prefill + input_batch.num_scheduled_tokens
         is_prompt_chunked = pos_after_step < prompt_lens
-        prefill_token_ids = self.req_states.prefill_token_ids
+        prefill_token_ids = self.req_states.prefill_token_ids.np
         query_start_loc = self.input_buffers.query_start_loc.np
         for i, req_id in enumerate(input_batch.req_ids):
             if not needs_prompt_logprobs[i]:
@@ -756,6 +756,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
             input_batch.idx_mapping,
             self.req_states.num_computed_tokens,
             self.req_states.last_sampled_tokens,
+            self.req_states.output_bin_counts,
             sampled_tokens,
             num_sampled,
             num_rejected,
@@ -785,7 +786,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         idx_mapping_np = input_batch.idx_mapping_np
         with async_barrier(self.spec_decode_event):
             self.input_buffers.next_prefill_tokens.np[:num_reqs] = (
-                self.req_states.prefill_token_ids[
+                self.req_states.prefill_token_ids.np[
                     idx_mapping_np,
                     self.req_states.num_computed_prefill_tokens[idx_mapping_np],
                 ]
@@ -896,7 +897,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                 # barrier to avoid race conditions.
                 pos = input_batch.positions[input_batch.logits_indices]
                 sampling_metadata = self.req_states.make_sampling_metadata(
-                    input_batch.idx_mapping_np, pos
+                    input_batch.idx_mapping, input_batch.idx_mapping_np, pos
                 )
                 if input_batch.num_draft_tokens > 0:
                     sampling_metadata = self.req_states.expand_sampling_metadata(

vllm/v1/worker/gpu/penalties.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import torch
+
+from vllm.triton_utils import tl, triton
+from vllm.v1.worker.gpu.states import SamplingMetadata
+
+
+@triton.jit
+def _penalties_kernel(
+    logits_ptr,
+    logits_stride,
+    repetition_penalty_ptr,
+    frequency_penalty_ptr,
+    presence_penalty_ptr,
+    idx_mapping_ptr,
+    prompt_bin_counts_ptr,
+    prompt_bin_counts_stride,
+    output_bin_counts_ptr,
+    output_bin_counts_stride,
+    vocab_size,
+    BLOCK_SIZE: tl.constexpr,
+):
+    batch_idx = tl.program_id(0)
+    rep_penalty = tl.load(repetition_penalty_ptr + batch_idx)
+    freq_penalty = tl.load(frequency_penalty_ptr + batch_idx)
+    pres_penalty = tl.load(presence_penalty_ptr + batch_idx)
+
+    use_rep_penalty = rep_penalty != 1.0
+    use_freq_penalty = freq_penalty != 0.0
+    use_pres_penalty = pres_penalty != 0.0
+    if not (use_rep_penalty or use_freq_penalty or use_pres_penalty):
+        # No penalties to apply. Early return.
+        return
+
+    block_idx = tl.program_id(1)
+    block = block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    mask = block < vocab_size
+    logits = tl.load(logits_ptr + batch_idx * logits_stride + block, mask=mask)
+    logits = logits.to(tl.float32)
+
+    req_state_idx = tl.load(idx_mapping_ptr + batch_idx)
+    output_bin_counts = tl.load(
+        output_bin_counts_ptr + req_state_idx * output_bin_counts_stride + block,
+        mask=mask,
+    )
+
+    # Apply repetition penalties.
+    if use_rep_penalty:
+        prompt_bin_counts = tl.load(
+            prompt_bin_counts_ptr + req_state_idx * prompt_bin_counts_stride + block,
+            mask=mask,
+        )
+        # If token appears in prompt or output, apply, otherwise use 1.0 for no-op.
+        scale = tl.where((prompt_bin_counts + output_bin_counts) > 0, rep_penalty, 1.0)
+        # If logits are positive, divide by penalty, otherwise multiply by penalty.
+        scale = tl.where(logits > 0, 1.0 / scale, scale)
+        logits *= scale
+
+    # Apply frequency penalties.
+    logits -= freq_penalty * output_bin_counts
+    # Apply presence penalties.
+    logits -= pres_penalty * (output_bin_counts > 0)
+    # Store back to logits.
+    tl.store(logits_ptr + batch_idx * logits_stride + block, logits, mask=mask)
+
+
+def apply_penalties(logits: torch.Tensor, sampling_metadata: SamplingMetadata) -> None:
+    num_reqs, vocab_size = logits.shape
+    BLOCK_SIZE = 8192
+    num_blocks = triton.cdiv(vocab_size, BLOCK_SIZE)
+    _penalties_kernel[(num_reqs, num_blocks)](
+        logits,
+        logits.stride(0),
+        sampling_metadata.repetition_penalty,
+        sampling_metadata.frequency_penalty,
+        sampling_metadata.presence_penalty,
+        sampling_metadata.idx_mapping,
+        sampling_metadata.prompt_bin_counts,
+        sampling_metadata.prompt_bin_counts.stride(0),
+        sampling_metadata.output_bin_counts,
+        sampling_metadata.output_bin_counts.stride(0),
+        vocab_size,
+        BLOCK_SIZE=BLOCK_SIZE,
+    )

vllm/v1/worker/gpu/sampler.py

@@ -8,6 +8,7 @@ from vllm.config.model import LogprobsMode
 from vllm.triton_utils import tl, triton
 from vllm.v1.outputs import LogprobsTensors, SamplerOutput
 from vllm.v1.sample.ops.topk_topp_sampler import apply_top_k_top_p
+from vllm.v1.worker.gpu.penalties import apply_penalties
 from vllm.v1.worker.gpu.states import SamplingMetadata
 
 
@@ -65,6 +66,8 @@ class Sampler:
         logits = apply_top_k_top_p(
             logits, sampling_metadata.top_k, sampling_metadata.top_p
         )
+        # Apply penalties in place.
+        apply_penalties(logits, sampling_metadata)
 
         sampled = gumbel_sample(
             logits,

vllm/v1/worker/gpu/states.py

@@ -8,6 +8,8 @@ import torch
 from vllm.lora.request import LoRARequest
 from vllm.sampling_params import SamplingParams
 from vllm.triton_utils import tl, triton
+from vllm.utils.platform_utils import is_uva_available
+from vllm.utils.torch_utils import get_cuda_view_from_cpu_tensor
 from vllm.v1.outputs import LogprobsTensors
 from vllm.v1.utils import CpuGpuBuffer
 
@@ -23,12 +25,21 @@ class SamplingMetadata:
     top_p: torch.Tensor | None
     top_k: torch.Tensor | None
 
+    repetition_penalty: torch.Tensor
+    frequency_penalty: torch.Tensor
+    presence_penalty: torch.Tensor
+
     seeds: torch.Tensor
     pos: torch.Tensor
 
     # None means no logprobs, 0 means sampled token logprobs only
     max_num_logprobs: int | None
 
+    # For penalties
+    idx_mapping: torch.Tensor
+    prompt_bin_counts: torch.Tensor
+    output_bin_counts: torch.Tensor
+
     @classmethod
     def make_dummy(
         cls,
@@ -44,17 +55,35 @@ class SamplingMetadata:
         # top_k = torch.full((num_reqs,), 20, dtype=torch.int32, device=device)
         top_p = None
         top_k = None
+        # NOTE(woosuk): We must set penalties to their default values to make sure
+        # the penalties kernel does not touch the placeholder bin_counts tensors.
+        repetition_penalty = torch.ones(num_reqs, dtype=torch.float32, device=device)
+        frequency_penalty = torch.zeros(num_reqs, dtype=torch.float32, device=device)
+        presence_penalty = torch.zeros(num_reqs, dtype=torch.float32, device=device)
         seeds = torch.zeros(num_reqs, dtype=torch.int64, device=device)
         pos = torch.zeros(num_reqs, dtype=torch.int64, device=device)
         max_num_logprobs = 20
 
+        idx_mapping = torch.arange(num_reqs, dtype=torch.int32, device=device)
+        # NOTE(woosuk): These are placeholder tensors to avoid None checks in the
+        # penalties kernel. We use 2 instead of 1 as vocab_size to avoid Triton
+        # specialization and re-compilation at runtime.
+        prompt_bin_counts = torch.zeros(num_reqs, 2, dtype=torch.int32, device=device)
+        output_bin_counts = torch.zeros(num_reqs, 2, dtype=torch.int32, device=device)
+
         return cls(
             temperature=temperature,
             top_p=top_p,
             top_k=top_k,
+            repetition_penalty=repetition_penalty,
+            frequency_penalty=frequency_penalty,
+            presence_penalty=presence_penalty,
             seeds=seeds,
             pos=pos,
             max_num_logprobs=max_num_logprobs,
+            idx_mapping=idx_mapping,
+            prompt_bin_counts=prompt_bin_counts,
+            output_bin_counts=output_bin_counts,
         )
 
 
@@ -83,9 +112,10 @@ class RequestState:
         self.extra_data: dict[str, ExtraData] = {}
 
         self.prompt_len = np.zeros(self.max_num_reqs, dtype=np.int32)
-        self.prefill_token_ids = np.zeros(
-            (self.max_num_reqs, self.max_model_len),
-            dtype=np.int32,
+        # NOTE(woosuk): This tensor can be extremely large (e.g., several GBs)
+        # depending on the configured max_num_reqs and max_model_len.
+        self.prefill_token_ids = UvaBuffer(
+            self.max_num_reqs, self.max_model_len, dtype=torch.int32
         )
         self.prefill_len = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
 
@@ -119,6 +149,9 @@ class RequestState:
         self.temperature = self._make_param(self.max_num_reqs, torch.float32)
         self.top_p = self._make_param(self.max_num_reqs, torch.float32)
         self.top_k = self._make_param(self.max_num_reqs, torch.int32)
+        self.repetition_penalty = self._make_param(self.max_num_reqs, torch.float32)
+        self.frequency_penalty = self._make_param(self.max_num_reqs, torch.float32)
+        self.presence_penalty = self._make_param(self.max_num_reqs, torch.float32)
         self.seeds = self._make_param(self.max_num_reqs, torch.int64)
 
         self.num_logprobs = np.empty(self.max_num_reqs, dtype=np.int32)
@@ -126,6 +159,16 @@ class RequestState:
         self.num_logprobs.fill(-1)
         self.needs_prompt_logprobs = np.zeros(self.max_num_reqs, dtype=bool)
 
+        # Statistics for penalties.
+        # TODO(woosuk): These tensors are rarely used but can be extremely large.
+        # Optimize the memory usage.
+        self.prompt_bin_counts = torch.zeros(
+            self.max_num_reqs, self.vocab_size, dtype=torch.int32, device=self.device
+        )
+        self.output_bin_counts = torch.zeros(
+            self.max_num_reqs, self.vocab_size, dtype=torch.int32, device=self.device
+        )
+
     def _make_param(self, size: int, dtype: torch.dtype) -> "Param":
         return Param(size, dtype=dtype, device=self.device, pin_memory=self.pin_memory)
 
@@ -159,7 +202,7 @@ class RequestState:
             f"prefill_len {prefill_len} < prompt_len {prompt_len}"
         )
         self.prefill_len.np[req_idx] = prefill_len
-        self.prefill_token_ids[req_idx, :prefill_len] = prefill_token_ids
+        self.prefill_token_ids.np[req_idx, :prefill_len] = prefill_token_ids
 
         self.num_computed_prefill_tokens[req_idx] = num_computed_tokens
         # FIXME(woosuk): This triggers a GPU operation whenever adding a new request.
@@ -178,6 +221,18 @@ class RequestState:
         else:
             top_k = self.vocab_size
         self.top_k.np[req_idx] = top_k
+        self.repetition_penalty.np[req_idx] = sampling_params.repetition_penalty
+        self.frequency_penalty.np[req_idx] = sampling_params.frequency_penalty
+        self.presence_penalty.np[req_idx] = sampling_params.presence_penalty
+
+        if use_penalty(sampling_params):
+            bincount(
+                self.prefill_token_ids.gpu[req_idx],
+                prefill_len,
+                prompt_len,
+                self.prompt_bin_counts[req_idx],
+                self.output_bin_counts[req_idx],
+            )
 
         if sampling_params.seed is not None:
             seed = sampling_params.seed
@@ -206,24 +261,32 @@ class RequestState:
 
     def make_sampling_metadata(
         self,
-        idx_mapping: np.ndarray,
+        idx_mapping: torch.Tensor,
+        idx_mapping_np: np.ndarray,
         pos: torch.Tensor,
     ) -> SamplingMetadata:
-        temperature = self.temperature.np[idx_mapping]
+        temperature = self.temperature.np[idx_mapping_np]
         temperature = self.temperature.copy_np_to_gpu(temperature)
 
-        top_p = self.top_p.np[idx_mapping]
+        top_p = self.top_p.np[idx_mapping_np]
         no_top_p = np.all(top_p == 1.0)
         top_p = self.top_p.copy_np_to_gpu(top_p) if not no_top_p else None
 
-        top_k = self.top_k.np[idx_mapping]
+        top_k = self.top_k.np[idx_mapping_np]
         no_top_k = np.all(top_k == self.vocab_size)
         top_k = self.top_k.copy_np_to_gpu(top_k) if not no_top_k else None
 
-        seeds = self.seeds.np[idx_mapping]
+        rep_penalty = self.repetition_penalty.np[idx_mapping_np]
+        rep_penalty = self.repetition_penalty.copy_np_to_gpu(rep_penalty)
+        freq_penalty = self.frequency_penalty.np[idx_mapping_np]
+        freq_penalty = self.frequency_penalty.copy_np_to_gpu(freq_penalty)
+        pres_penalty = self.presence_penalty.np[idx_mapping_np]
+        pres_penalty = self.presence_penalty.copy_np_to_gpu(pres_penalty)
+
+        seeds = self.seeds.np[idx_mapping_np]
         seeds = self.seeds.copy_np_to_gpu(seeds)
 
-        num_logprobs = self.num_logprobs[idx_mapping]
+        num_logprobs = self.num_logprobs[idx_mapping_np]
         max_num_logprobs: int | None = int(np.max(num_logprobs))
         if max_num_logprobs == -1:
             max_num_logprobs = None
@@ -232,9 +295,15 @@ class RequestState:
             temperature=temperature,
             top_p=top_p,
             top_k=top_k,
+            repetition_penalty=rep_penalty,
+            frequency_penalty=freq_penalty,
+            presence_penalty=pres_penalty,
             seeds=seeds,
             pos=pos,
             max_num_logprobs=max_num_logprobs,
+            idx_mapping=idx_mapping,
+            prompt_bin_counts=self.prompt_bin_counts,
+            output_bin_counts=self.output_bin_counts,
         )
 
     def expand_sampling_metadata(
@@ -294,6 +363,14 @@ class ExtraData:
     in_progress_prompt_logprobs: list[LogprobsTensors] = field(default_factory=list)
 
 
+class UvaBuffer:
+    def __init__(self, *size: int | torch.SymInt, dtype: torch.dtype):
+        assert is_uva_available()
+        self.cpu = torch.zeros(*size, dtype=dtype, device="cpu", pin_memory=True)
+        self.np = self.cpu.numpy()
+        self.gpu = get_cuda_view_from_cpu_tensor(self.cpu)
+
+
 # NOTE(woosuk): Re-compilation can happen at runtime since top_p and top_k can be None.
 @triton.jit
 def _expand_sampling_metadata_kernel(
@@ -304,6 +381,12 @@ def _expand_sampling_metadata_kernel(
     top_k_ptr,
     expanded_top_k_ptr,
     seeds_ptr,
+    rep_penalty_ptr,
+    expanded_rep_penalty_ptr,
+    freq_penalty_ptr,
+    expanded_freq_penalty_ptr,
+    pres_penalty_ptr,
+    expanded_pres_penalty_ptr,
     expanded_seeds_ptr,
     cu_num_logits_ptr,
     BLOCK_SIZE: tl.constexpr,
@@ -327,6 +410,15 @@ def _expand_sampling_metadata_kernel(
         top_k = tl.load(top_k_ptr + req_idx)
         tl.store(expanded_top_k_ptr + start_idx + block, top_k, mask=mask)
 
+    rep_penalty = tl.load(rep_penalty_ptr + req_idx)
+    tl.store(expanded_rep_penalty_ptr + start_idx + block, rep_penalty, mask=mask)
+
+    freq_penalty = tl.load(freq_penalty_ptr + req_idx)
+    tl.store(expanded_freq_penalty_ptr + start_idx + block, freq_penalty, mask=mask)
+
+    pres_penalty = tl.load(pres_penalty_ptr + req_idx)
+    tl.store(expanded_pres_penalty_ptr + start_idx + block, pres_penalty, mask=mask)
+
     seed = tl.load(seeds_ptr + req_idx)
     tl.store(expanded_seeds_ptr + start_idx + block, seed, mask=mask)
 
@@ -341,6 +433,9 @@ def expand_sampling_metadata(
     expanded_temp = create_empty(sampling_metadata.temperature)
     expanded_top_p = create_empty(sampling_metadata.top_p)
     expanded_top_k = create_empty(sampling_metadata.top_k)
+    expanded_repetition_penalty = create_empty(sampling_metadata.repetition_penalty)
+    expanded_frequency_penalty = create_empty(sampling_metadata.frequency_penalty)
+    expanded_presence_penalty = create_empty(sampling_metadata.presence_penalty)
     expanded_seeds = create_empty(sampling_metadata.seeds)
 
     num_reqs = cu_num_logits.shape[0] - 1
@@ -351,6 +446,12 @@ def expand_sampling_metadata(
         expanded_top_p,
         sampling_metadata.top_k,
         expanded_top_k,
+        sampling_metadata.repetition_penalty,
+        expanded_repetition_penalty,
+        sampling_metadata.frequency_penalty,
+        expanded_frequency_penalty,
+        sampling_metadata.presence_penalty,
+        expanded_presence_penalty,
         sampling_metadata.seeds,
         expanded_seeds,
         cu_num_logits,
@@ -361,6 +462,67 @@ def expand_sampling_metadata(
         top_p=expanded_top_p,
         top_k=expanded_top_k,
         seeds=expanded_seeds,
+        repetition_penalty=expanded_repetition_penalty,
+        frequency_penalty=expanded_frequency_penalty,
+        presence_penalty=expanded_presence_penalty,
         pos=sampling_metadata.pos,
         max_num_logprobs=sampling_metadata.max_num_logprobs,
+        # TODO(woosuk): Support penalties with spec decoding.
+        idx_mapping=sampling_metadata.idx_mapping,
+        prompt_bin_counts=sampling_metadata.prompt_bin_counts,
+        output_bin_counts=sampling_metadata.output_bin_counts,
+    )
+
+
+def use_penalty(sampling_params: SamplingParams) -> bool:
+    return (
+        sampling_params.repetition_penalty != 1.0
+        or sampling_params.frequency_penalty != 0.0
+        or sampling_params.presence_penalty != 0.0
+    )
+
+
+@triton.jit(do_not_specialize=["prefill_len", "prompt_len"])
+def _bincount_kernel(
+    prefill_token_ids_ptr,
+    prefill_len,
+    prompt_len,
+    prompt_bin_counts_ptr,
+    output_bin_counts_ptr,
+    BLOCK_SIZE: tl.constexpr,
+):
+    block_idx = tl.program_id(0)
+    if block_idx * BLOCK_SIZE >= prefill_len:
+        return
+
+    block = block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    if block_idx * BLOCK_SIZE < prompt_len:
+        mask = block < prompt_len
+        prefill_tokens = tl.load(prefill_token_ids_ptr + block, mask=mask)
+        tl.atomic_add(prompt_bin_counts_ptr + prefill_tokens, 1, mask=mask)
+    if (block_idx + 1) * BLOCK_SIZE >= prompt_len:
+        mask = block < prefill_len
+        mask &= block >= prompt_len
+        prefill_tokens = tl.load(prefill_token_ids_ptr + block, mask=mask)
+        tl.atomic_add(output_bin_counts_ptr + prefill_tokens, 1, mask=mask)
+
+
+def bincount(
+    prefill_token_ids: torch.Tensor,
+    prefill_len: int,
+    prompt_len: int,
+    prompt_bin_counts: torch.Tensor,
+    output_bin_counts: torch.Tensor,
+) -> None:
+    prompt_bin_counts.zero_()
+    output_bin_counts.zero_()
+    BLOCK_SIZE = 1024
+    num_blocks = triton.cdiv(prefill_len, BLOCK_SIZE)
+    _bincount_kernel[(num_blocks,)](
+        prefill_token_ids,
+        prefill_len,
+        prompt_len,
+        prompt_bin_counts,
+        output_bin_counts,
+        BLOCK_SIZE=BLOCK_SIZE,
     )