Fix docs for clip max new tokens (#9082)

docs/advanced_features/pd_disaggregation.md

@@ -67,7 +67,6 @@ Please be aware that this setting will cause prefill instances to take a longer
 | **`SGLANG_DISAGGREGATION_HEARTBEAT_INTERVAL`** | Interval (seconds) between health checks to prefill bootstrap servers | `5.0` |
 | **`SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE`** | Consecutive heartbeat failures before marking prefill server offline | `2` |
 | **`SGLANG_DISAGGREGATION_WAITING_TIMEOUT`** | Timeout (seconds) for receiving KV Cache after request initialization | `300` |
-| **`SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION`** | Clip request param "max_tokens" to pre_allocate | `4096` |
 
 If a greater mean TTFT is acceptable, you can `export SGLANG_DISAGGREGATION_WAITING_TIMEOUT=600` (10 minutes) to relax the timeout condition.
 

docs/references/environment_variables.md

@@ -45,7 +45,7 @@ SGLang supports various environment variables that can be used to configure its
 | Environment Variable | Description | Default Value |
 | --- | --- | --- |
 | `SGLANG_DEBUG_MEMORY_POOL` | Enable memory pool debugging | `false` |
-| `SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION` | Clip max new tokens estimation for memory planning | Not set |
+| `SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION` | Clip max new tokens estimation for memory planning | `4096` |
 | `SGLANG_DETOKENIZER_MAX_STATES` | Maximum states for detokenizer | Default value based on system |
 | `SGL_DISABLE_TP_MEMORY_INBALANCE_CHECK` | Disable checks for memory imbalance across Tensor Parallel ranks | Not set (defaults to enabled check) |
 

python/sglang/srt/disaggregation/decode.py

@@ -59,9 +59,7 @@ if TYPE_CHECKING:
     from sglang.srt.managers.schedule_batch import Req
     from sglang.srt.managers.scheduler import Scheduler
 
-DECODE_CLIP_MAX_NEW_TOKEN = get_int_env_var(
+CLIP_MAX_NEW_TOKEN = get_int_env_var("SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION", 4096)
-    "SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION", 4096
-)
 
 
 class DecodeReqToTokenPool:
@@ -390,7 +388,7 @@ class DecodePreallocQueue:
                     origin_input_len
                     + min(
                         decode_req.req.sampling_params.max_new_tokens,
-                        DECODE_CLIP_MAX_NEW_TOKEN,
+                        CLIP_MAX_NEW_TOKEN,
                     )
                     - retractable_tokens,
                 )
@@ -440,7 +438,7 @@ class DecodePreallocQueue:
         need_space_for_single_req = (
             max(
                 [
-                    min(x.sampling_params.max_new_tokens, DECODE_CLIP_MAX_NEW_TOKEN)
+                    min(x.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKEN)
                     + len(x.origin_input_ids)
                     - retractable_tokens
                     for x in self.scheduler.running_batch.reqs

python/sglang/srt/managers/schedule_policy.py

@@ -36,7 +36,7 @@ if TYPE_CHECKING:
 # This can prevent the server from being too conservative.
 # Note that this only clips the estimation in the scheduler but does not change the stop
 # condition. The request can still generate tokens until it hits the unclipped max_new_tokens.
-CLIP_MAX_NEW_TOKENS_ESTIMATION = int(
+CLIP_MAX_NEW_TOKENS = int(
     os.environ.get("SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION", "4096")
 )
 
@@ -305,7 +305,7 @@ class PrefillAdder:
                 [
                     min(
                         (r.sampling_params.max_new_tokens - len(r.output_ids)),
-                        CLIP_MAX_NEW_TOKENS_ESTIMATION,
+                        CLIP_MAX_NEW_TOKENS,
                     )
                     * self.new_token_ratio
                     for r in running_batch.reqs
@@ -388,7 +388,7 @@ class PrefillAdder:
             0,
             req.extend_input_len,
             (
-                min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS_ESTIMATION)
+                min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS)
                 if not truncated
                 else 0
             ),
@@ -477,7 +477,7 @@ class PrefillAdder:
             self._update_prefill_budget(
                 0,
                 req.extend_input_len,
-                min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS_ESTIMATION),
+                min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS),
             )
         else:
             if self.rem_chunk_tokens == 0:
@@ -499,7 +499,7 @@ class PrefillAdder:
             return self.add_one_req_ignore_eos(req, has_chunked_req)
 
         total_tokens = req.extend_input_len + min(
-            req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS_ESTIMATION
+            req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS
         )
 
         # adjusting the input_tokens based on host_hit_length and page_size
@@ -544,7 +544,7 @@ class PrefillAdder:
                     input_tokens,
                     min(
                         req.sampling_params.max_new_tokens,
-                        CLIP_MAX_NEW_TOKENS_ESTIMATION,
+                        CLIP_MAX_NEW_TOKENS,
                     ),
                 )
             else: