Slightly improve the sampler to skip unnecessary steps (#6956)

python/sglang/srt/layers/sampler.py

@@ -5,7 +5,7 @@ import torch
 import torch.distributed as dist
 from torch import nn
 
-from sglang.srt.distributed import get_tensor_model_parallel_group
+from sglang.srt.distributed import get_tp_group
 from sglang.srt.layers.dp_attention import get_attention_tp_group
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
 from sglang.srt.managers.schedule_batch import global_server_args_dict
@@ -30,7 +30,7 @@ class Sampler(nn.Module):
     def __init__(self):
         super().__init__()
         self.use_nan_detection = global_server_args_dict["enable_nan_detection"]
-        self.tp_sync_group = get_tensor_model_parallel_group().device_group
+        self.tp_sync_group = get_tp_group().device_group
 
         if global_server_args_dict["enable_dp_attention"]:
             self.tp_sync_group = get_attention_tp_group().device_group
@@ -59,7 +59,7 @@ class Sampler(nn.Module):
 
         # Apply the custom logit processors if registered in the sampling info.
         if sampling_info.has_custom_logit_processor:
-            self._apply_custom_logit_processor(logits, sampling_info)
+            apply_custom_logit_processor(logits, sampling_info)
 
         if self.use_nan_detection and torch.any(torch.isnan(logits)):
             logger.warning("Detected errors during sampling! NaN in the logits.")
@@ -81,16 +81,8 @@ class Sampler(nn.Module):
             probs = logits
             del logits
 
+            if True:  # Keep this redundant check to simplify some internal code sync
                 if global_server_args_dict["sampling_backend"] == "flashinfer":
-                if return_logprob:
-                    # NOTE: the top_p_renorm_prob from flashinfer has numerical problems,
-                    # https://github.com/flashinfer-ai/flashinfer/issues/708
-                    # so we use the torch implementation.
-                    # NOTE: OpenAI's logprobs is independent of top-p, we use the
-                    # same rule.
-                    logprobs = torch.log(probs).clamp(min=torch.finfo(probs.dtype).min)
-
-                max_top_k_round, batch_size = 32, probs.shape[0]
                     if sampling_info.need_min_p_sampling:
                         probs = top_k_renorm_prob(probs, sampling_info.top_ks)
                         probs = top_p_renorm_prob(probs, sampling_info.top_ps)
@@ -98,16 +90,13 @@ class Sampler(nn.Module):
                             probs, sampling_info.min_ps
                         )
                     else:
-                    # Check Nan will throw exception, only check when crash_on_warnings is True
-                    check_nan = self.use_nan_detection and crash_on_warnings()
                         batch_next_token_ids = top_k_top_p_sampling_from_probs(
-                        probs.contiguous(),
+                            probs,
                             sampling_info.top_ks,
                             sampling_info.top_ps,
                             filter_apply_order="joint",
-                        check_nan=check_nan,
+                            check_nan=self.use_nan_detection,
                         )
-
                 elif global_server_args_dict["sampling_backend"] == "pytorch":
                     # A slower fallback implementation with torch native operations.
                     batch_next_token_ids = top_k_top_p_min_p_sampling_from_probs_torch(
@@ -117,14 +106,15 @@ class Sampler(nn.Module):
                         sampling_info.min_ps,
                         sampling_info.need_min_p_sampling,
                     )
-
-                if return_logprob:
-                    logprobs = torch.log(probs).clamp(min=torch.finfo(probs.dtype).min)
                 else:
                     raise ValueError(
                         f"Invalid sampling backend: {global_server_args_dict['sampling_backend']}"
                     )
 
+            if return_logprob:
+                # clamp to avoid -inf
+                logprobs = torch.log(probs).clamp(min=torch.finfo(probs.dtype).min)
+
         # Attach logprobs to logits_output (in-place modification)
         if return_logprob:
             if any(x > 0 for x in top_logprobs_nums):
@@ -160,39 +150,6 @@ class Sampler(nn.Module):
 
         return batch_next_token_ids
 
-    def _apply_custom_logit_processor(
-        self, logits: torch.Tensor, sampling_batch_info: SamplingBatchInfo
-    ):
-        """Apply custom logit processors to the logits.
-        This function will modify the logits in-place."""
-
-        assert logits.shape[0] == len(sampling_batch_info), (
-            f"The batch size of logits ({logits.shape[0]}) does not match the batch size of "
-            f"sampling_batch_info ({len(sampling_batch_info)})"
-        )
-
-        for _, (
-            processor,
-            batch_mask,
-        ) in sampling_batch_info.custom_logit_processor.items():
-            # Get the batch indices that need to be processed
-            batch_indices = batch_mask.nonzero(as_tuple=True)[0]
-
-            assert batch_mask.shape[0] == len(sampling_batch_info), (
-                f"The number of batch mask ({batch_mask.shape[0]}) does not match the number of "
-                f"sampling_batch_info ({len(sampling_batch_info)})"
-            )
-
-            # Apply the processor to the logits
-            logits[batch_mask] = processor(
-                logits[batch_mask],
-                [sampling_batch_info.custom_params[i] for i in batch_indices],
-            )
-
-            logger.debug(
-                f"Custom logit processor {processor.__class__.__name__} is applied."
-            )
-
 
 def top_k_top_p_min_p_sampling_from_probs_torch(
     probs: torch.Tensor,
@@ -221,6 +178,14 @@ def top_k_top_p_min_p_sampling_from_probs_torch(
     return batch_next_token_ids
 
 
+def sampling_from_probs_torch(probs: torch.Tensor):
+    """A sampling implementation with native pytorch operations, without
+    top-k, top-p, or min-p filtering."""
+    sampled_index = torch.multinomial(probs, num_samples=1)
+    batch_next_token_ids = sampled_index.view(-1).to(torch.int32)
+    return batch_next_token_ids
+
+
 def top_p_normalize_probs_torch(
     probs: torch.Tensor,
     top_ps: torch.Tensor,
@@ -259,3 +224,44 @@ def get_token_ids_logprobs(logprobs: torch.Tensor, token_ids_logprobs: List[List
             output_token_ids_logprobs_idx.append([])
 
     return output_token_ids_logprobs_val, output_token_ids_logprobs_idx
+
+
+def apply_custom_logit_processor(
+    logits: torch.Tensor,
+    sampling_batch_info: SamplingBatchInfo,
+    num_tokens_in_batch: int = 1,
+):
+    """Apply custom logit processors to the logits.
+    This function will modify the logits in-place.
+    num_tokens_in_batch is needed to support spec decoding, where each batch can contain multiple
+    tokens. By default, we assume each batch contains only 1 token.
+    """
+
+    assert logits.shape[0] == len(sampling_batch_info) * num_tokens_in_batch, (
+        f"The batch size of logits ({logits.shape[0]}) does not match the batch size of "
+        f"sampling_batch_info ({len(sampling_batch_info)}) x num_tokens_in_batch "
+        f"({num_tokens_in_batch})"
+    )
+
+    for _, (
+        processor,
+        batch_mask,
+    ) in sampling_batch_info.custom_logit_processor.items():
+        # Get the batch indices that need to be processed
+        batch_indices = batch_mask.nonzero(as_tuple=True)[0]
+
+        assert batch_mask.shape[0] == len(sampling_batch_info), (
+            f"The number of batch mask ({batch_mask.shape[0]}) does not match the number of "
+            f"sampling_batch_info ({len(sampling_batch_info)})"
+        )
+        batch_mask = torch.repeat_interleave(batch_mask, num_tokens_in_batch)
+
+        # Apply the processor to the logits
+        logits[batch_mask] = processor(
+            logits[batch_mask],
+            [sampling_batch_info.custom_params[i] for i in batch_indices],
+        )
+
+        logger.debug(
+            f"Custom logit processor {processor.__class__.__name__} is applied."
+        )

python/sglang/srt/managers/tokenizer_manager.py

@@ -852,7 +852,7 @@ class TokenizerManager:
             obj.load_format = self.server_args.load_format
         logger.info("Start update_weights. Load format=%s", obj.load_format)
 
-        if True:
+        if True:  # Keep this redundant check to simplify some internal code sync
             # Hold the lock if it is not async. This means that weight sync
             # cannot run while requests are in progress.
             async with self.model_update_lock.writer_lock:

python/sglang/srt/models/llama.py

@@ -17,7 +17,7 @@
 """Inference-only LLaMA model compatible with HuggingFace weights."""
 
 import logging
-from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Union
+from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
 
 import torch
 from torch import nn

python/sglang/srt/sampling/sampling_batch_info.py

@@ -9,10 +9,12 @@ import torch
 
 import sglang.srt.sampling.penaltylib as penaltylib
 from sglang.srt.sampling.custom_logit_processor import CustomLogitProcessor
+from sglang.srt.sampling.sampling_params import TOP_K_ALL
 
 if TYPE_CHECKING:
     from sglang.srt.managers.schedule_batch import ScheduleBatch
 
+
 logger = logging.getLogger(__name__)
 
 
@@ -27,6 +29,12 @@ class SamplingBatchInfo:
     # Whether all requests use greedy sampling
     is_all_greedy: bool
 
+    # Whether any requests use top_p sampling
+    need_top_p_sampling: bool
+
+    # Whether any requests use top_k sampling
+    need_top_k_sampling: bool
+
     # Whether any request needs min_p sampling
     need_min_p_sampling: bool
 
@@ -133,6 +141,8 @@ class SamplingBatchInfo:
             top_ks=top_ks,
             min_ps=min_ps,
             is_all_greedy=all(r.sampling_params.top_k <= 1 for r in reqs),
+            need_top_p_sampling=any(r.sampling_params.top_p != 1.0 for r in reqs),
+            need_top_k_sampling=any(r.sampling_params.top_k != TOP_K_ALL for r in reqs),
             need_min_p_sampling=any(r.sampling_params.min_p > 0 for r in reqs),
             vocab_size=vocab_size,
             penalizer_orchestrator=penalizer_orchestrator,
@@ -167,7 +177,7 @@ class SamplingBatchInfo:
 
         # Apply the mask
         for i, grammar in enumerate(self.grammars):
-            if grammar and not grammar.finished:
+            if grammar and not grammar.finished and not grammar.is_terminated():
                 grammar.fill_vocab_mask(self.vocab_mask, i)
 
         # Move the mask to the device if needed
@@ -308,4 +318,6 @@ class SamplingBatchInfo:
             setattr(self, item, torch.cat([self_val, other_val]))
 
         self.is_all_greedy &= other.is_all_greedy
+        self.need_top_p_sampling |= other.need_top_p_sampling
+        self.need_top_k_sampling |= other.need_top_k_sampling
         self.need_min_p_sampling |= other.need_min_p_sampling

python/sglang/srt/sampling/sampling_params.py

@@ -16,6 +16,7 @@
 from typing import Any, Dict, List, Optional, Union
 
 _SAMPLING_EPS = 1e-6
+TOP_K_ALL = 1 << 30
 
 
 class SamplingParams:
@@ -84,7 +85,7 @@ class SamplingParams:
             self.temperature = 1.0
             self.top_k = 1
         if self.top_k == -1:
-            self.top_k = 1 << 30  # whole vocabulary
+            self.top_k = TOP_K_ALL  # whole vocabulary
 
     def verify(self):
         if self.temperature < 0.0:

python/sglang/srt/two_batch_overlap.py

 import dataclasses
 import logging
-from typing import TYPE_CHECKING, Dict, List, Optional, Sequence
+from typing import Dict, List, Optional, Sequence
 
 import torch
 
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
 from sglang.srt.layers.communicator import (
     CommunicateContext,
-    CommunicateSimpleFn,
     CommunicateSummableTensorPairFn,
     ScatterMode,
 )
-from sglang.srt.layers.dp_attention import get_attention_tp_size
 from sglang.srt.layers.moe.ep_moe.token_dispatcher import DeepEPDispatcher
 from sglang.srt.layers.quantization.deep_gemm import configure_deep_gemm_num_sms
 from sglang.srt.managers.schedule_batch import global_server_args_dict
@@ -20,9 +18,6 @@ from sglang.srt.operations import execute_operations, execute_overlapped_operati
 from sglang.srt.operations_strategy import OperationsStrategy
 from sglang.srt.utils import BumpAllocator, DeepEPMode, get_bool_env_var
 
-if TYPE_CHECKING:
-    from sglang.srt.model_executor.cuda_graph_runner import CudaGraphRunner
-
 _tbo_debug = get_bool_env_var("SGLANG_TBO_DEBUG")
 
 logger = logging.getLogger(__name__)
@@ -46,7 +41,7 @@ def compute_split_seq_index(
         assert num_tokens == 0
         return 0
     else:
-        raise NotImplementedError
+        raise NotImplementedError()
 
 
 def _split_array_by_half_sum(arr: Sequence[int]) -> int:

python/sglang/srt/utils.py

@@ -1928,16 +1928,18 @@ def next_power_of_2(n: int):
 setattr(triton, "next_power_of_2", next_power_of_2)
 
 
-@contextmanager
-def empty_context(*args, **kwargs):
-    try:
-        # Setup code goes here
-        yield
-    finally:
-        # Cleanup code goes here
+class EmptyContextManager:
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
         pass
 
 
+def empty_context(*args, **kwargs):
+    return EmptyContextManager()
+
+
 def add_prefix(name: str, prefix: str) -> str:
     """Add a weight path prefix to a module name.