Fix bugs in sampler with CUDA graph / torch.compile (#1306)

python/sglang/srt/layers/sampler.py

@@ -1,6 +1,6 @@
 import dataclasses
 import logging
-from typing import Union
+from typing import Tuple, Union
 
 import torch
 from flashinfer.sampling import (
@@ -9,6 +9,7 @@ from flashinfer.sampling import (
     top_k_top_p_sampling_from_probs,
     top_p_renorm_prob,
 )
+from torch.library import custom_op as torch_custom_op
 from vllm.model_executor.custom_op import CustomOp
 
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
@@ -30,6 +31,9 @@ class SampleOutput:
 class Sampler(CustomOp):
     def __init__(self):
         super().__init__()
+        # FIXME: torch.multinomial has too many bugs
+        self.forward_native = self.forward_cuda
+        self.is_torch_compile = False
 
     def _apply_penalties(self, logits: torch.Tensor, sampling_info: SamplingBatchInfo):
         # min-token, presence, frequency
@@ -46,16 +50,11 @@ class Sampler(CustomOp):
 
         return logits
 
-    def _get_probs(
+    def _get_probs(self, logits: torch.Tensor, sampling_info: SamplingBatchInfo):
-        self,
-        logits: torch.Tensor,
-        sampling_info: SamplingBatchInfo,
-        is_torch_compile: bool = False,
-    ):
         # Post process logits
         logits = logits.contiguous()
         logits.div_(sampling_info.temperatures)
-        if is_torch_compile:
+        if self.is_torch_compile:
             # FIXME: Temporary workaround for unknown bugs in torch.compile
             logits.add_(0)
 
@@ -91,7 +90,7 @@ class Sampler(CustomOp):
                     probs, uniform_samples, sampling_info.min_ps
                 )
             else:
-                batch_next_token_ids, success = top_k_top_p_sampling_from_probs(
+                batch_next_token_ids, success = flashinfer_top_k_top_p(
                     probs, uniform_samples, sampling_info.top_ks, sampling_info.top_ps
                 )
         else:
@@ -110,7 +109,7 @@ class Sampler(CustomOp):
         if isinstance(logits, LogitsProcessorOutput):
             logits = logits.next_token_logits
 
-        probs = self._get_probs(logits, sampling_info, is_torch_compile=True)
+        probs = self._get_probs(logits, sampling_info)
 
         batch_next_token_ids, success = top_k_top_p_min_p_sampling_from_probs_torch(
             probs, sampling_info.top_ks, sampling_info.top_ps, sampling_info.min_ps
@@ -119,6 +118,31 @@ class Sampler(CustomOp):
         return SampleOutput(success, probs, batch_next_token_ids)
 
 
+@torch_custom_op("my_lib::flashinfer_top_k_top_p", mutates_args={})
+def flashinfer_top_k_top_p(
+    probs: torch.Tensor,
+    uniform_samples: torch.Tensor,
+    top_ks: torch.Tensor,
+    top_ps: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    # NOTE: we do not use min_p neither in CUDA nor in torch.compile
+    return top_k_top_p_sampling_from_probs(probs, uniform_samples, top_ks, top_ps)
+
+
+@flashinfer_top_k_top_p.register_fake
+def _(
+    probs: torch.Tensor,
+    uniform_samples: torch.Tensor,
+    top_ks: torch.Tensor,
+    top_ps: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    bs = probs.shape[0]
+    return (
+        torch.ones(bs, dtype=torch.bool, device=probs.device),
+        torch.zeros(bs, dtype=torch.int32, device=probs.device),
+    )
+
+
 def top_k_top_p_min_p_sampling_from_probs_torch(
     probs: torch.Tensor,
     top_ks: torch.Tensor,

python/sglang/srt/model_executor/cuda_graph_runner.py

@@ -46,8 +46,10 @@ def _to_torch(model: torch.nn.Module, reverse: bool = False):
         if isinstance(sub, CustomOp):
             if reverse:
                 sub._forward_method = sub.forward_cuda
+                setattr(sub, "is_torch_compile", False)
             else:
                 sub._forward_method = sub.forward_native
+                setattr(sub, "is_torch_compile", True)
         if isinstance(sub, torch.nn.Module):
             _to_torch(sub, reverse)
 

python/sglang/srt/model_executor/model_runner.py

@@ -523,7 +523,7 @@ class ModelRunner:
         if (
             self.cuda_graph_runner
             and self.cuda_graph_runner.can_run(len(batch.reqs))
-            and not batch.sampling_info.has_bias()
+            and batch.sampling_info.can_run_in_cuda_graph()
         ):
             return self.cuda_graph_runner.replay(batch)
 

python/sglang/srt/sampling/sampling_batch_info.py

@@ -34,12 +34,14 @@ class SamplingBatchInfo:
     linear_penalties: torch.Tensor = None
     scaling_penalties: torch.Tensor = None
 
-    def has_bias(self):
+    def can_run_in_cuda_graph(self):
+        # Vocab bias and min_ps are not supported in CUDA graph
         return (
-            self.logit_bias is not None
+            self.logit_bias is None
-            or self.vocab_mask is not None
+            and self.vocab_mask is None
-            or self.linear_penalties is not None
+            and self.linear_penalties is None
-            or self.scaling_penalties is not None
+            and self.scaling_penalties is None
+            and not self.need_min_p_sampling
         )
 
     @classmethod
@@ -48,35 +50,29 @@ class SamplingBatchInfo:
         ret.temperatures = torch.ones((max_bs, 1), dtype=torch.float, device="cuda")
         ret.top_ps = torch.ones((max_bs,), dtype=torch.float, device="cuda")
         ret.top_ks = torch.ones((max_bs,), dtype=torch.int, device="cuda")
-        ret.min_ps = torch.zeros((max_bs,), dtype=torch.float, device="cuda")
         return ret
 
     def __getitem__(self, key):
         if isinstance(key, slice):
-            # NOTE: We do not use cuda graph when there is bias tensors
+            # NOTE:This method is only used in CUDA graph
-            assert not self.has_bias()
+            assert self.can_run_in_cuda_graph()
             return SamplingBatchInfo(
                 vocab_size=self.vocab_size,
                 temperatures=self.temperatures[key],
                 top_ps=self.top_ps[key],
                 top_ks=self.top_ks[key],
-                min_ps=self.min_ps[key],
-                need_min_p_sampling=self.need_min_p_sampling,
             )
         else:
             raise NotImplementedError
 
     def inplace_assign(self, bs: int, other: SamplingBatchInfo):
-        # NOTE: We do not use cuda graph when there is bias tensors
+        # NOTE:This method is only used in CUDA graph
-        assert not self.has_bias()
+        assert self.can_run_in_cuda_graph()
 
         self.vocab_size = other.vocab_size
-        self.need_min_p_sampling = other.need_min_p_sampling
-
         self.temperatures[:bs] = other.temperatures
         self.top_ps[:bs] = other.top_ps
         self.top_ks[:bs] = other.top_ks
-        self.min_ps[:bs] = other.min_ps
 
     @classmethod
     def from_schedule_batch(cls, batch: ScheduleBatch, vocab_size: int):