patches: serve-404-served-names — the model-not-found 404 lists the served names

PATCHES.md

@@ -46,6 +46,7 @@ build by name instead of landing by guess. Regenerate a file with `bash scripts/
 | qwen3_5-embed-quant | fix | pass `quant_config` to the token embedding (main model and MTP module) | none yet | 0.28.0 | upstream PR |
 | qwen3_5-mtp-draft-vocab | feature | vocab-truncated draft head for MTP | none | 0.28.0 | upstreamed |
 | sampler-small-topk-fast-softmax | feature | sort-free top-k/top-p for small k, multi-block row softmax; registers `VLLM_DRAFT_TOPK_TOPP` and `VLLM_DRAFT_TEMP_SCALE` (both read once at import) | none | 0.28.0 | upstreamed or superseded |
+| serve-404-served-names | fix | the model-not-found 404 lists the served names (`Served models: ...`) so a misnamed model is a one-read response body | vllm #58025 | 0.28.0 | upstream PR |
 | spec-decode-attn | feature | split-KV verify attention on FLASH_ATTN with query-row tiling; registers `VLLM_SPEC_DECODE_ATTN`, `VLLM_SPEC_DECODE_ATTN_QMAX`, `VLLM_SPEC_ATTN_BLOCK_M` (#114) | none | 0.28.0 | upstreamed |
 | spec-decode-int4-kv-mq3d | feature | multi-query 3D int4 verify path | none | 0.28.0 | rides with int4-kv-per-token-head |
 | spec-decode-int8-kv | feature | split-KV verify attention over an int8 per-token-head cache | none | 0.28.0 | rides with spec-decode-attn |

patches/series

@@ -51,4 +51,8 @@ engine-stall-sentinel.patch
 sse-keep-alive.patch
 int4-mq3d-envs.patch
 triton-spec-attn-fp8-kv.patch
+<<<<<<< HEAD
 bench-probe-errors.patch
+=======
+serve-404-served-names.patch
+>>>>>>> 48e63ae (patches: serve-404-served-names — the model-not-found 404 lists the served names)

patches/serve-404-served-names.patch

@@ -0,0 +1,31 @@
+The model-not-found 404 now lists the served names.
+
+"The model `X` does not exist." sends the reader to the server log to find
+out what WOULD have been accepted. /v1/models already knows; the 404 body
+now carries the same list, so a misnamed model becomes a one-read response
+body instead of an incident:
+
+    The model `X` does not exist. Served models: qwen3.8-27b.
+
+--- exported from cpuchip/vllm 1c5d490 (serve-404-served-names); regenerate with scripts/export-patch.sh, do not edit ---
+
+diff --git a/entrypoints/serve/engine/serving.py b/entrypoints/serve/engine/serving.py
+index a31e47d..ada47d6 100644
+--- a/entrypoints/serve/engine/serving.py
++++ b/entrypoints/serve/engine/serving.py
+@@ -60,8 +60,14 @@
+             ):
+                 error_response = load_result
+
++        served_names = ", ".join(
++            model.name for model in self.models.base_model_paths
++        )
+         return error_response or self.create_error_response(
+-            message=f"The model `{request.model}` does not exist.",
++            message=(
++                f"The model `{request.model}` does not exist. "
++                f"Served models: {served_names}."
++            ),
+             err_type="NotFoundError",
+             status_code=HTTPStatus.NOT_FOUND,
+             param="model",