Move sampling logits to float32 (#773)

python/sglang/srt/layers/logits_processor.py

@@ -136,7 +136,7 @@ class LogitsProcessor(nn.Module):
         last_logits = torch.matmul(last_hidden, weight.T)
         if self.tp_size > 1:
             last_logits = tensor_model_parallel_all_gather(last_logits)
-        last_logits = last_logits[:, : self.config.vocab_size]
+        last_logits = last_logits[:, : self.config.vocab_size].float()
 
         if hasattr(self.config, "final_logit_softcapping"):
             last_logits /= self.config.final_logit_softcapping
@@ -161,9 +161,9 @@ class LogitsProcessor(nn.Module):
                 all_logits = torch.matmul(hidden_states, weight.T)
                 if self.tp_size > 1:
                     all_logits = tensor_model_parallel_all_gather(all_logits)
-                all_logits = all_logits[:, : self.config.vocab_size]
+                all_logits = all_logits[:, : self.config.vocab_size].float()
 
-            all_logprobs = all_logits.float()
+            all_logprobs = all_logits
             del all_logits
             all_logprobs[:] = torch.nn.functional.log_softmax(all_logprobs, dim=-1)
 

python/sglang/srt/managers/controller/infer_batch.py

@@ -687,13 +687,21 @@ class Batch:
         # TODO(lmzheng): apply penalty
         probs = torch.softmax(logits, dim=-1)
 
+        if True:
             max_top_k_round, batch_size = 32, probs.shape[0]
-        uniform_samples = torch.rand((max_top_k_round, batch_size), device=probs.device)
+            uniform_samples = torch.rand(
+                (max_top_k_round, batch_size), device=probs.device
+            )
             batch_next_token_ids, success = top_k_top_p_sampling_from_probs(
                 probs, uniform_samples, self.top_ks, self.top_ps
             )
+        else:
+            # Here we provide a slower fallback implementation.
+            batch_next_token_ids, success = top_k_top_p_sampling_from_probs_torch(
+                probs, self.top_ks, self.top_ps
+            )
 
-        if torch.any(~success):
+        if not torch.all(success):
             warnings.warn("Sampling failed, fallback to top_k=1 strategy")
             probs = probs.masked_fill(torch.isnan(probs), 0.0)
             argmax_ids = torch.argmax(probs, dim=-1)
@@ -933,3 +941,29 @@ def init_triton_args(forward_mode, seq_lens, prefix_lens):
         max_extend_len = int(torch.max(extend_seq_lens))
 
     return max_seq_len, max_extend_len, start_loc, prefix_lens
+
+
+def top_k_top_p_sampling_from_probs_torch(
+    probs: torch.Tensor, top_ks: torch.Tensor, top_ps: torch.Tensor
+):
+    """A top-k and top-k sampling implementation with native pytorch operations."""
+    probs_sort, probs_idx = probs.sort(dim=-1, descending=True)
+    probs_sum = torch.cumsum(probs_sort, dim=-1)
+    probs_sort[(probs_sum - probs_sort) > top_ps.view(-1, 1)] = 0.0
+    probs_sort[
+        torch.arange(0, probs.shape[-1], device=probs.device).view(1, -1)
+        >= top_ks.view(-1, 1)
+    ] = 0.0
+    probs_sort.div_(probs_sort.max(dim=-1, keepdim=True)[0])
+    try:
+        sampled_index = torch.multinomial(probs_sort, num_samples=1)
+    except RuntimeError:
+        batch_next_token_ids = torch.zeros(
+            (probs_sort.shape[0],), dtype=torch.int64, device=probs.device
+        )
+        success = torch.zeros(probs.shape[0], dtype=torch.bool, device=probs.device)
+        return batch_next_token_ids, success
+
+    batch_next_token_ids = torch.gather(probs_idx, dim=1, index=sampled_index).view(-1)
+    success = torch.ones(probs.shape[0], dtype=torch.bool, device=probs.device)
+    return batch_next_token_ids, success