[BugFix] Add compatibility guard for unsupported CLI params on Ascend NPU (vllm-project#10602)

## What this PR does

Add silent reset handling in `_fix_incompatible_config` for four
upstream CLI parameters that are incompatible with Ascend NPU:

| Parameter | Config Section | Reset Value | Reason |
|-----------|---------------|-------------|--------|
| `--calculate-kv-scales` | Cache Config | `False` | Deprecated
upstream; not supported on Ascend |
| `--compilation-config.use_inductor_graph_partition` | Compilation
Config | `False` | Ascend forces `use_inductor=False`, making this
irrelevant |

## User-facing change

Users who pass `--calculate-kv-scales`, `--compilation-config
'{"use_inductor_graph_partition": true}'` on Ascend NPU will now see a
warning log and the parameter will be silently reset to its safe
default, rather than potentially causing errors or undefined behavior.

## How was this patch tested

- Code review of the `_fix_incompatible_config` method to verify the
pattern matches existing silent-reset entries.


- vLLM version: v0.22.1
- vLLM main:
vllm-project/vllm@967c5c3

---------

Signed-off-by: underfituu <hzhucong@163.com>

Author: underfituu

vllm_ascend/platform.py

@@ -1049,6 +1049,13 @@ def _fix_incompatible_config(vllm_config: VllmConfig) -> None:
                 )
                 vllm_config.cache_config.cpu_kvcache_space_bytes = None
 
+            if getattr(vllm_config.cache_config, "calculate_kv_scales", False):
+                logger.warning(
+                    "Parameter is not supported on Ascend NPU. "
+                    "parameter=calculate_kv_scales, action: resetting to False."
+                )
+                vllm_config.cache_config.calculate_kv_scales = False
+
         # ==================== 3. MultiModal Config ====================
         multimodal_config = getattr(model_config, "multimodal_config", None) if model_config else None
         if multimodal_config:
@@ -1224,6 +1231,15 @@ def _fix_incompatible_config(vllm_config: VllmConfig) -> None:
                 )
                 vllm_config.parallel_config.ubatch_size = 0
 
+        # ==================== 10. Compilation Config ====================
+        if vllm_config.compilation_config:
+            if getattr(vllm_config.compilation_config, "use_inductor_graph_partition", False):
+                logger.warning(
+                    "Parameter is not supported on Ascend NPU (use_inductor is False). "
+                    "parameter=use_inductor_graph_partition, action: resetting to False."
+                )
+                vllm_config.compilation_config.use_inductor_graph_partition = False
+
     @classmethod
     def use_custom_op_collectives(cls) -> bool:
         return True