[Config] Update default _max_num_batched_tokens from 8192 to 16384 (#51726)

Signed-off-by: yewentao256 <zhyanwentao@126.com>

tests/v1/engine/test_engine_args.py

@@ -5,7 +5,6 @@ from argparse import ArgumentError
 
 import pytest
 
-from vllm.config import VllmConfig
 from vllm.engine.arg_utils import EngineArgs
 from vllm.usage.usage_lib import UsageContext
 from vllm.utils.argparse_utils import FlexibleArgumentParser
@@ -64,34 +63,6 @@ def test_prefix_caching_xxhash_from_cli():
     assert vllm_config.cache_config.prefix_caching_hash_algo == "xxhash_cbor"
 
 
-def test_defaults_with_usage_context():
-    engine_args = EngineArgs(model="facebook/opt-125m")
-    vllm_config: VllmConfig = engine_args.create_engine_config(UsageContext.LLM_CLASS)
-
-    from vllm.platforms import current_platform
-    from vllm.utils.mem_constants import GiB_bytes
-
-    device_memory = current_platform.get_device_total_memory()
-    device_name = current_platform.get_device_name().lower()
-    if device_memory >= 70 * GiB_bytes and "a100" not in device_name:
-        # For GPUs like H100, H200, and MI300x with >= 70GB memory
-        default_llm_tokens = 16384
-        default_server_tokens = 8192
-        default_max_num_seqs = 1024
-    else:
-        default_llm_tokens = 8192
-        default_server_tokens = 2048
-        default_max_num_seqs = 256
-
-    assert vllm_config.scheduler_config.max_num_seqs == default_max_num_seqs
-    assert vllm_config.scheduler_config.max_num_batched_tokens == default_llm_tokens  # noqa: E501
-
-    engine_args = EngineArgs(model="facebook/opt-125m")
-    vllm_config = engine_args.create_engine_config(UsageContext.OPENAI_API_SERVER)
-    assert vllm_config.scheduler_config.max_num_seqs == default_max_num_seqs
-    assert vllm_config.scheduler_config.max_num_batched_tokens == default_server_tokens  # noqa: E501
-
-
 def test_mm_prefix_lm_raises_batched_tokens_floor():
     """Verify that prefix-LM multimodal models auto-raise
     max_num_batched_tokens to fit at least one multimodal item.

vllm/engine/arg_utils.py

@@ -2566,8 +2566,18 @@ class EngineArgs:
         # NOTE(Kuntai): Setting large `max_num_batched_tokens` for A100 reduces
         # throughput, see PR #17885 for more details.
         # So here we do an extra device name check to prevent such regression.
-        if device_memory >= 70 * GiB_bytes and "a100" not in device_name:
-            # For GPUs like H100 and MI300x, use larger default values.
+        if device_memory >= 160 * GiB_bytes:
+            # for GPUs like B200/B300 with >= 160GB memory, use the largest defaults
+            default_max_num_batched_tokens = {
+                UsageContext.LLM_CLASS: 16384,
+                UsageContext.OPENAI_API_SERVER: 16384,
+            }
+            default_max_num_seqs = {
+                UsageContext.LLM_CLASS: 1024,
+                UsageContext.OPENAI_API_SERVER: 1024,
+            }
+        elif device_memory >= 70 * GiB_bytes and "a100" not in device_name:
+            # For GPUs like H100 and H200, use larger offline defaults.
             default_max_num_batched_tokens = {
                 UsageContext.LLM_CLASS: 16384,
                 UsageContext.OPENAI_API_SERVER: 8192,