[Core] Use flashinfer sampling kernel when available (vllm-project#7137)

Co-authored-by: Michael Goin <michael@neuralmagic.com>

.buildkite/test-pipeline.yaml

@@ -192,7 +192,9 @@ steps:
   - vllm/model_executor/layers
   - vllm/sampling_metadata.py
   - tests/samplers
-  command: pytest -v -s samplers
+  commands:
+    - pytest -v -s samplers
+    - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
 
 - label: LogitsProcessor Test # 5min
   mirror_hardwares: [amd]

Dockerfile

@@ -194,7 +194,7 @@ RUN --mount=type=bind,from=mamba-builder,src=/usr/src/mamba,target=/usr/src/mamb
     python3 -m pip install /usr/src/mamba/*.whl --no-cache-dir
 
 RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.3/flashinfer-0.1.3+cu121torch2.4-cp310-cp310-linux_x86_64.whl
+    python3 -m pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.4/flashinfer-0.1.4+cu121torch2.4-cp310-cp310-linux_x86_64.whl
 #################### vLLM installation IMAGE ####################
 
 

tests/samplers/test_sampler.py

@@ -8,6 +8,7 @@
 import torch
 from transformers import GenerationConfig, GenerationMixin
 
+import vllm.envs as envs
 from vllm.model_executor.layers.sampler import Sampler
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.model_executor.utils import set_random_seed
@@ -634,7 +635,10 @@ def mock_sample(probs, *args, **kwargs):
         return ([[prob.topk(1, dim=-1).indices.tolist(), [0]]
                  for prob in probs], None)
 
-    with patch("vllm.model_executor.layers.sampler._sample", mock_sample):
+    # top-k and top-p is only calculated when flashinfer kernel is not available
+    with patch("vllm.model_executor.layers.sampler._sample", mock_sample), \
+         patch("vllm.model_executor.layers.sampler."
+               "flashinfer_top_k_top_p_sampling", None):
         sampler(logits=fake_logits, sampling_metadata=sampling_metadata)
 
     assert sample_probs is not None
@@ -645,6 +649,37 @@ def mock_sample(probs, *args, **kwargs):
     assert torch.equal(hf_probs.eq(0), sample_probs.eq(0))
 
 
+@pytest.mark.parametrize("seed", RANDOM_SEEDS)
+@pytest.mark.parametrize("device", CUDA_DEVICES)
+def test_flashinfer_fallback(seed: int, device: str):
+    if not envs.VLLM_USE_FLASHINFER_SAMPLER:
+        pytest.skip("Flashinfer sampler is disabled")
+
+    set_random_seed(seed)
+    torch.set_default_device(device)
+    batch_size = random.randint(1, 256)
+    _, fake_logits, sampler = _prepare_test(batch_size)
+
+    def failing_flashinfer_sampling(*_args, **_kwargs):
+        return None, torch.zeros(batch_size, device=device, dtype=torch.int32)
+
+    sampling_params = SamplingParams(
+        temperature=1.0,
+        n=random.randint(1, 10),
+        seed=random.randint(0, 10000),
+    )
+    sampler_output = _do_sample(batch_size, fake_logits, sampler,
+                                sampling_params, device)
+
+    with patch(
+            "vllm.model_executor.layers.sampler."
+            "flashinfer_top_k_top_p_sampling", failing_flashinfer_sampling):
+        fallback_sampler_output = _do_sample(batch_size, fake_logits, sampler,
+                                             sampling_params, device)
+
+    assert sampler_output == fallback_sampler_output
+
+
 @pytest.mark.parametrize("device", CUDA_DEVICES)
 def test_sampler_repetition_penalty_mixed(device: str):
 

vllm/envs.py

@@ -30,6 +30,7 @@
     VLLM_LOGGING_CONFIG_PATH: Optional[str] = None
     VLLM_TRACE_FUNCTION: int = 0
     VLLM_ATTENTION_BACKEND: Optional[str] = None
+    VLLM_USE_FLASHINFER_SAMPLER: bool = False
     VLLM_PP_LAYER_PARTITION: Optional[str] = None
     VLLM_CPU_KVCACHE_SPACE: int = 0
     VLLM_CPU_OMP_THREADS_BIND: str = ""
@@ -256,6 +257,10 @@ def get_default_config_root():
     "VLLM_ATTENTION_BACKEND":
     lambda: os.getenv("VLLM_ATTENTION_BACKEND", None),
 
+    # If set, vllm will use flashinfer sampler
+    "VLLM_USE_FLASHINFER_SAMPLER":
+    lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_SAMPLER", "0"))),
+
     # Pipeline stage partition strategy
     "VLLM_PP_LAYER_PARTITION":
     lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),

vllm/model_executor/layers/sampler.py

@@ -1,5 +1,7 @@
 """A layer that samples the next tokens from the model's outputs."""
 import itertools
+import warnings
+from importlib.util import find_spec
 from math import inf
 from typing import Dict, List, Optional, Tuple
 
@@ -11,6 +13,7 @@
 if HAS_TRITON:
     from vllm.model_executor.layers.ops.sample import sample as sample_triton
 
+import vllm.envs as envs
 from vllm.model_executor.sampling_metadata import (SamplingMetadata,
                                                    SamplingTensors,
                                                    SequenceGroupToSample)
@@ -19,6 +22,16 @@
                            PromptLogprobs, SampleLogprobs, SamplerOutput,
                            SequenceOutput)
 
+if envs.VLLM_USE_FLASHINFER_SAMPLER and find_spec("flashinfer"):
+    import flashinfer.sampling
+    # yapf: disable
+    from flashinfer.sampling import (
+        top_k_top_p_sampling_from_probs as flashinfer_top_k_top_p_sampling)
+
+    # yapf: enable
+else:
+    flashinfer_top_k_top_p_sampling = None
+
 # (num_token_ids, num_parent_ids) per sequence group.
 SampleResultType = List[Tuple[List[int], List[int]]]
 
@@ -123,7 +136,7 @@ def forward(
         logits = logits.to(torch.float)
         logits.div_(sampling_tensors.temperatures.unsqueeze(dim=1))
 
-        if do_top_p_top_k:
+        if do_top_p_top_k and flashinfer_top_k_top_p_sampling is None:
             logits = _apply_top_k_top_p(logits, sampling_tensors.top_ps,
                                         sampling_tensors.top_ks)
 
@@ -476,32 +489,65 @@ def _multinomial(
     seq_groups: Optional[List[SequenceGroupToSample]] = None,
 ) -> torch.Tensor:
     if num_samples > 1:
-        # This is equivalent to torch.repeat_interleaved (which also
-        # forces a GPU<->CPU sync).
-        # This allows us to do sampling with replacement by creating
-        # num_samples copies of each row in the tensor, and then
-        # batch sampling the resulting tensor.
-        probs = probs[:, None, :].expand(probs.shape[0], num_samples,
-                                         probs.shape[1]).contiguous().view(
-                                             -1, probs.shape[1])
+        probs = probs.repeat_interleave(num_samples, dim=0)
     q = torch.empty_like(probs)
     if seq_groups is None:
         q.exponential_()
     else:
         sample_idx = 0
         for seq_group in seq_groups:
             seq_ids = seq_group.seq_ids
-            next_sample_idx = sample_idx + len(seq_ids) * num_samples
-            q[sample_idx:next_sample_idx].exponential_(
-                generator=seq_group.generator)
-            sample_idx = next_sample_idx
+            stride = len(seq_ids) * num_samples
+            assert seq_group.generator is not None
+            q[sample_idx:sample_idx +
+              stride].exponential_(generator=seq_group.generator)
+            sample_idx += stride
     return probs.div_(q).argmax(dim=1).view(-1, num_samples)
 
 
+def _top_k_top_p_multinomial_with_flashinfer(
+        probs: torch.Tensor, top_ks: torch.Tensor, top_ps: torch.Tensor,
+        num_samples: int, seq_groups: Optional[List[SequenceGroupToSample]]):
+    max_top_k_round = 32
+    if num_samples > 1:
+        probs = probs.repeat_interleave(num_samples, dim=0)
+        top_ks = top_ks.repeat_interleave(num_samples)
+        top_ps = top_ps.repeat_interleave(num_samples)
+    batch_size = probs.shape[0]
+    uniform_samples = torch.empty((max_top_k_round, batch_size),
+                                  device=probs.device)
+    if seq_groups is None:
+        uniform_samples.uniform_()
+    else:
+        sample_idx = 0
+        for seq_group in seq_groups:
+            seq_ids = seq_group.seq_ids
+            stride = len(seq_ids) * num_samples
+            assert seq_group.generator is not None
+            uniform_samples[:, sample_idx:sample_idx +
+                            stride].uniform_(generator=seq_group.generator)
+            sample_idx += stride
+    batch_next_token_ids, success = flashinfer_top_k_top_p_sampling(
+        probs,
+        uniform_samples,
+        top_ks,
+        top_ps,
+    )
+    if not success.all():
+        warnings.warn("FlashInfer rejection sampling failed, fallback.",
+                      stacklevel=1)
+        probs = flashinfer.sampling.top_k_renorm_prob(probs, top_ks)
+        probs = flashinfer.sampling.top_p_renorm_prob(probs, top_ps)
+        batch_next_token_ids = flashinfer.sampling.sampling_from_probs(
+            probs, uniform_samples[0])
+    return batch_next_token_ids.view(-1, num_samples)
+
+
 def _sample_with_torch(
     probs: torch.Tensor,
     logprobs: torch.Tensor,
     sampling_metadata: SamplingMetadata,
+    sampling_tensors: SamplingTensors,
     include_gpu_probs_tensor: bool,
     modify_greedy_probs: bool,
 ) -> Tuple[SampleResultType, Optional[torch.Tensor]]:
@@ -564,18 +610,28 @@ def _sample_with_torch(
                     sampling_params = seq_group.sampling_params
                     max_best_of_in_batch = max(max_best_of_in_batch,
                                                sampling_params.best_of)
-            seeded_args = {} if sampling_type == SamplingType.RANDOM else {
-                "seq_groups": seq_groups,
-            }
-
-            multinomial_samples[sampling_type] = _multinomial(
-                probs[long_sample_indices], max_best_of_in_batch,
-                **seeded_args)
+            seq_groups_arg = (None if sampling_type == SamplingType.RANDOM else
+                              seq_groups)
+
+            if flashinfer_top_k_top_p_sampling is not None:
+                multinomial_samples[
+                    sampling_type] = _top_k_top_p_multinomial_with_flashinfer(
+                        probs[long_sample_indices],
+                        sampling_tensors.top_ks[long_sample_indices],
+                        sampling_tensors.top_ps[long_sample_indices],
+                        max_best_of_in_batch,
+                        seq_groups_arg,
+                    )
+            else:
+                multinomial_samples[sampling_type] = _multinomial(
+                    probs[long_sample_indices],
+                    max_best_of_in_batch,
+                    seq_groups=seq_groups_arg)
 
             if sampled_token_ids_tensor is not None:
                 # Store sampled tokens in output tensor.
-                sampled_token_ids_tensor[
-                    long_sample_indices] = multinomial_samples[sampling_type]
+                sampled_token_ids_tensor[long_sample_indices] = \
+                    multinomial_samples[sampling_type].to(torch.long)
 
         elif sampling_type == SamplingType.BEAM:
             beam_search_logprobs = logprobs[sample_indices]
@@ -693,9 +749,12 @@ def _sample_with_triton_kernel(
 
 
 def _sample(
-    probs: torch.Tensor, logprobs: torch.Tensor,
-    sampling_metadata: SamplingMetadata, sampling_tensors: SamplingTensors,
-    include_gpu_probs_tensor: bool, modify_greedy_probs: bool
+    probs: torch.Tensor,
+    logprobs: torch.Tensor,
+    sampling_metadata: SamplingMetadata,
+    sampling_tensors: SamplingTensors,
+    include_gpu_probs_tensor: bool,
+    modify_greedy_probs: bool,
 ) -> Tuple[SampleResultType, Optional[torch.Tensor]]:
     """
     Args:
@@ -713,6 +772,7 @@ def _sample(
         probs,
         logprobs,
         sampling_metadata,
+        sampling_tensors,
         include_gpu_probs_tensor=include_gpu_probs_tensor,
         modify_greedy_probs=modify_greedy_probs,
     )