Sync code and test CI; rename some env vars (#11686)

python/sglang/srt/speculative/eagle_worker.py

@@ -64,7 +64,7 @@ if is_cuda():
     from sgl_kernel import segment_packbits
 
 logger = logging.getLogger(__name__)
-RETURN_ORIGINAL_LOGPROB = get_bool_env_var("RETURN_ORIGINAL_LOGPROB")
+SGLANG_RETURN_ORIGINAL_LOGPROB = get_bool_env_var("SGLANG_RETURN_ORIGINAL_LOGPROB")
 
 
 @contextmanager
@@ -741,7 +741,7 @@ class EAGLEWorker(TpModelWorker):
         # acceptance indices are the indices in a "flattened" batch.
         # dividing it to num_draft_tokens will yield the actual batch index.
         temperatures = temperatures[accepted_indices // num_draft_tokens]
-        if RETURN_ORIGINAL_LOGPROB:
+        if SGLANG_RETURN_ORIGINAL_LOGPROB:
             logprobs = torch.nn.functional.log_softmax(
                 logits_output.next_token_logits, dim=-1
             )

python/sglang/srt/speculative/spec_info.py

@@ -1,5 +1,6 @@
 from abc import ABC, abstractmethod
 from enum import IntEnum, auto
+from functools import lru_cache
 from typing import List, Tuple
 
 from sglang.srt.managers.schedule_batch import ModelWorkerBatch
@@ -27,6 +28,7 @@ class SpeculativeAlgorithm(IntEnum):
     def is_ngram(self):
         return self == SpeculativeAlgorithm.NGRAM
 
+    @lru_cache(maxsize=None)
     @staticmethod
     def from_string(name: str):
         name_map = {

python/sglang/srt/speculative/standalone_worker.py

@@ -15,7 +15,7 @@ if is_cuda():
     from sgl_kernel import segment_packbits
 
 logger = logging.getLogger(__name__)
-RETURN_ORIGINAL_LOGPROB = get_bool_env_var("RETURN_ORIGINAL_LOGPROB")
+SGLANG_RETURN_ORIGINAL_LOGPROB = get_bool_env_var("SGLANG_RETURN_ORIGINAL_LOGPROB")
 
 
 @contextmanager