feat(models): add the vllm_native dialect (RFC #25's seventh)

docs/guide/configuration.md

@@ -93,7 +93,8 @@ root model (top level or `args`) and on an inline `grader` / `extractor`, on
 every dialect. A derived model (`base:`) uses its root's key.
 
 With neither, the `openai_*` dialects read `OPENAI_API_KEY`; the others send
-no key, so an external SGLang server started with `--api-key` needs one here.
+no key, so an external SGLang or vLLM server started with `--api-key` needs
+one here.
 
 ```yaml
 models:
@@ -182,7 +183,7 @@ one entry per rollout, in `texts` order, with the same three fields.
 
 | Dialect | `n: 1` | `n > 1` |
 | --- | --- | --- |
-| `sglang_native`, `openai_completions`, `openai_chat` | yes | yes |
+| `sglang_native`, `vllm_native`, `openai_completions`, `openai_chat` | yes | yes |
 | `google_genai` | yes | refused |
 | `openai_responses` | yes | refused (no `n`) |
 | `anthropic_messages` | refused | refused |
@@ -194,8 +195,10 @@ so changing them on a resume reruns inference.
 With `echo` (`alogprobs(text, echo=True, logprobs=N)` or `infer_args: {echo:
 true, top_logprobs: N}`), each input token's top-N lands in
 `ModelOutput.input_top_logprobs` (`{}` for the first); `top_logprobs` keeps
-the generated tokens'. Served by `sglang_native` and `openai_completions`.
-The reply grows by N entries per input token.
+the generated tokens'. Served by `sglang_native`, `vllm_native` and
+`openai_completions`. The reply grows by N entries per input token.
+
+`sglang_native` and `vllm_native` also fill token ids.
 
 ## Model Capabilities (`extra`)
 

docs/guide/infer.md

@@ -257,6 +257,22 @@ task scores input: SGLang's `/v1/completions` cannot. `/generate` has no
 per-request `seed`, so an explicit one is refused; a managed engine pins
 `--random-seed` instead.
 
+A vLLM `gen` model keeps `openai_completions` unless it sets `dialect:
+vllm_native`, which posts to the same `/v1/completions` with vLLM's own
+fields (measured on 0.26.0):
+
+| Need | `openai_completions` | `vllm_native` |
+| --- | --- | --- |
+| Input scoring | `echo`, split at `usage.prompt_tokens` | `prompt_logprobs`, checked against `prompt_token_ids` |
+| `texts` of a scoring call | prompt + generation | generation only |
+| Token ids | none | input and sampled (`return_token_ids`) |
+| Structured output | refused | `structured_outputs` |
+| Unknown options | `extra_body` | refused, except `min_p`, `repetition_penalty`, `stop_token_ids` |
+
+It sends one non-streaming POST and offers no reasoning or tool parsers
+(they act on chat requests). `suffix` and `max_tokens: 0` are refused before
+I/O. Input top-k needs `--max-logprobs` at least the breadth.
+
 ### Model lifecycles
 
 A model `sieval run` launches may bound its engine by stage barriers

sieval/core/models/dialect_registry.py

@@ -54,6 +54,10 @@
     CAPABILITY_DECISIONS as SGLANG_NATIVE_CAPABILITY_DECISIONS,
 )
 from .dialects.sglang_native import SglangNativeDialect
+from .dialects.vllm_native import (
+    CAPABILITY_DECISIONS as VLLM_NATIVE_CAPABILITY_DECISIONS,
+)
+from .dialects.vllm_native import VllmNativeDialect
 from .ir import (
     ChatInput,
     ChatMessage,
@@ -378,6 +382,10 @@ def _bind_sglang_native(connection: Any, requested_model_id: str) -> Dialect:
     return SglangNativeDialect(connection, requested_model_id)
 
 
+def _bind_vllm_native(connection: Any, requested_model_id: str) -> Dialect:
+    return VllmNativeDialect(connection, requested_model_id)
+
+
 def _binder(
     function: Callable[[Any, str], Dialect],
     decisions: Mapping[str, DialectCapabilityDecision],
@@ -433,6 +441,13 @@ def _binder(
                 SGLANG_NATIVE_CAPABILITY_DECISIONS,
             ),
         ),
+        "vllm_native": _binder(
+            _bind_vllm_native,
+            cast(
+                Mapping[str, DialectCapabilityDecision],
+                VLLM_NATIVE_CAPABILITY_DECISIONS,
+            ),
+        ),
     }
 )
 
@@ -504,7 +519,9 @@ def _registered_decisions(
         "vllm_native": _spec(
             "vllm_native",
             "async_http_json",
-            request_seed_support=RequestSeedSupport.RESERVED,
+            decisions=_registered_decisions("vllm_native"),
+            # Measured on 0.26.0: equal seeds give equal samples at temperature 1.
+            request_seed_support=RequestSeedSupport.SUPPORTED,
             input_kinds=("completion",),
             input_modalities=("text",),
         ),
@@ -628,8 +645,6 @@ def bind_dialect(
 
 
 def _not_implemented_message(dialect_id: str) -> str:
-    if dialect_id == "vllm_native":
-        return "dialect 'vllm_native' is reserved and has no executable binder"
     return f"dialect {dialect_id!r} is reserved and has no executable binder"