refactor : merge max_prompt_len into max_seq_len, replace assert with raise

- Engine/Scheduler/TaskManager: merge max_prompt_len into max_seq_len
- train.py: replace bare assert with ValueError/FileNotFoundError
- server.py: add --max_seq_len CLI option
- engine.py: remove dead page_size param

scripts/tools/benchmark.py

@@ -47,7 +47,6 @@ class GenerationBenchmark:
             tokenizer=None,
             max_batch_size=256,
             max_seq_len=config.max_position_embeddings,
-            max_prompt_len=config.max_position_embeddings,
         )
 
     def run_prefill_benchmark(

scripts/tools/generate.py

@@ -38,7 +38,6 @@ def processor(
         tokenizer=tokenizer,
         max_batch_size=batch_size * num_samples,
         max_seq_len=cache_len,
-        max_prompt_len=cache_len,
     )
 
     print(f"Reading {input_json_file} ...")

scripts/tools/server.py

@@ -31,7 +31,15 @@ _DTYPES = ["bfloat16", "float16", "float32"]
     default=16,
     help="Maximum batch size for continuous batching.",
 )
-def server_command(host, port, reload, param_path, device, dtype, max_batch_size):
+@click.option(
+    "--max_seq_len",
+    type=int,
+    default=None,
+    help="Maximum sequence length (KV cache size + prompt truncation). Uses model config if not set.",
+)
+def server_command(
+    host, port, reload, param_path, device, dtype, max_batch_size, max_seq_len
+):
     """Launch inference server (OpenAI-compatible API)."""
     dtype_map = {
         "bfloat16": torch.bfloat16,
@@ -51,6 +59,7 @@ def server_command(host, port, reload, param_path, device, dtype, max_batch_size
         dtype=dtype_map[dtype],
         param_path=Path(param_path),
         max_batch_size=max_batch_size,
+        max_seq_len=max_seq_len,
     )
 
 

scripts/tools/train.py

@@ -402,15 +402,20 @@ def train(
     decay_steps: int,
     **kwargs,
 ):
-    assert train_type in [
+    if train_type not in [
         "seq",
         "sft",
         "dpo",
         "grpo",
         "online_grpo",
         "online_dpo",
-    ]
-    assert os.path.exists(param_path)
+    ]:
+        raise ValueError(
+            f"Invalid train_type '{train_type}'. "
+            f"Must be one of: seq, sft, dpo, grpo, online_grpo, online_dpo"
+        )
+    if not os.path.exists(param_path):
+        raise FileNotFoundError(f"Model directory not found: {param_path}")
     if nprocs > 1 and parallel_mode == "none":
         raise ValueError(
             "--nprocs > 1 requires --parallel_mode to be 'ddp', 'fsdp', or 'fsdp2'"