[Refactor] MLP weight prefetch to consistency with MoE Model's prefetching in terms of code and usage (#6442)

### What this PR does / why we need it?
Refactor MLP weight prefetch to consistency with MoE Model's prefetching
in terms of code and usage.
Environments VLLM_ASCEND_ENABLE_PREFETCH_MLP,
VLLM_ASCEND_MLP_DOWN_PREFETCH_SIZE and
VLLM_ASCEND_MLP_GATE_UP_PREFETCH_SIZE is removed, usage as following:

--additional-config '{"weight_prefetch_config": { "enabled": true,
"prefetch_ratio": {"mlp": { "gate_up": 1.0, "down": 1.0} }}}'

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.14.1
- vLLM main:
dc917cceb8

---------

Signed-off-by: leo-pony <nengjunma@outlook.com>

tests/e2e/nightly/single_node/models/test_qwen3_32b_int8.py

@@ -83,7 +83,6 @@ async def test_models(model: str, mode: str, tp_size: int) -> None:
         "TASK_QUEUE_ENABLE": "1",
         "HCCL_OP_EXPANSION_MODE": "AIV",
         "VLLM_ASCEND_ENABLE_FLASHCOMM": "1",
-        "VLLM_ASCEND_ENABLE_PREFETCH_MLP": "1"
     }
     compilation_config = {
         "cudagraph_mode":
@@ -98,7 +97,8 @@ async def test_models(model: str, mode: str, tp_size: int) -> None:
         str(port), "--max-model-len", "40960", "--max-num-batched-tokens",
         "40960", "--block-size", "128", "--trust-remote-code",
         "--reasoning-parser", "qwen3", "--gpu-memory-utilization", "0.9",
-        "--async-scheduling"
+        "--async-scheduling", "--additional-config",
+        '{"weight_prefetch_config":{"enabled":true}}',
     ]
     if mode == "single":
         server_args.append("--enforce-eager")

tests/e2e/nightly/single_node/models/test_qwen3_32b_int8_a3_feature_stack3.py

@@ -72,7 +72,6 @@ async def test_models(model: str, tp_size: int) -> None:
         "OMP_PROC_BIND": "false",
         "VLLM_ASCEND_ENABLE_TOPK_OPTIMIZE": "1",
         "VLLM_ASCEND_ENABLE_FLASHCOMM": "1",
-        "VLLM_ASCEND_ENABLE_PREFETCH_MLP": "1"
     }
     server_args = [
         "--quantization", "ascend", "--tensor-parallel-size",
@@ -82,7 +81,8 @@ async def test_models(model: str, tp_size: int) -> None:
         "0.9", "--block-size", "128", "--max-num-seqs", "256",
         "--enforce-eager", "--max-model-len", "35840",
         "--max-num-batched-tokens", "35840", "--additional-config",
-        '{"enable_weight_nz_layout":true}', "--compilation-config",
+        '{"enable_weight_nz_layout":true, "weight_prefetch_config":{"enabled": true}}',
+        "--compilation-config",
         '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[1,8,24,48,60]}'
     ]
     with RemoteOpenAIServer(model,

tests/e2e/nightly/single_node/models/test_qwq_32b.py

@@ -75,8 +75,7 @@ async def test_models(model: str, mode: str, tp_size: int) -> None:
         "OMP_PROC_BIND": "false",
         "HCCL_OP_EXPANSION_MODE": "AIV",
         "VLLM_ASCEND_ENABLE_FLASHCOMM": "1",
-        "VLLM_ASCEND_ENABLE_DEBSE_OPTIMIZE": "1",
-        "VLLM_ASCEND_ENABLE_PREFETCH_MLP": "1"
+        "VLLM_ASCEND_ENABLE_DEBSE_OPTIMIZE": "1"
     }
     server_args = [
         "--tensor-parallel-size",
@@ -86,7 +85,7 @@ async def test_models(model: str, mode: str, tp_size: int) -> None:
         "--gpu-memory-utilization", "0.9", "--compilation_config",
         '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes": [1, 8, 24, 48, 60]}',
         "--reasoning-parser", "deepseek_r1", "--distributed_executor_backend",
-        "mp"
+        "mp", "--additional-config", '{"weight_prefetch_config":{"enabled":true}}'
     ]
     if mode == "single":
         server_args.remove("--compilation_config")