feat(RL): forward routed_experts_prompt_start via nvext

components/src/dynamo/vllm/handlers.py

@@ -753,6 +753,19 @@ def build_sampling_params(
         if value is not None and hasattr(sampling_params, key):
             setattr(sampling_params, key, value)
 
+    # routed_experts_prompt_start (RL capture offset) rides nvext, not the
+    # standard sampling_options (Dynamo's chat schema rejects unknown top-level
+    # fields). Apply it onto SamplingParams so vLLM trims the leading prompt rows
+    # from the returned routing engine-side (instead of the client trimming the
+    # full-sequence blob after it crosses the wire).
+    for source in _iter_nvext_sources(request):
+        reps_nvext = source.get("routed_experts_prompt_start")
+        if reps_nvext is not None and hasattr(
+            sampling_params, "routed_experts_prompt_start"
+        ):
+            sampling_params.routed_experts_prompt_start = reps_nvext
+            break
+
     # routed_experts_prompt_start (RL capture offset) must be a non-negative
     # int; reject bad client values so the worker emits a sane `start` instead
     # of a bogus offset the consumer cannot align (vLLM clamps the upper bound).

components/src/dynamo/vllm/tests/test_vllm_unit.py

@@ -900,6 +900,33 @@ def test_build_sampling_params_maps_max_thinking_tokens():
     assert sp.thinking_token_budget == 1024
 
 
+def test_build_sampling_params_applies_nvext_routed_experts_prompt_start():
+    """routed_experts_prompt_start rides nvext (not sampling_options) and is
+    applied onto SamplingParams so vLLM trims routing engine-side."""
+    import pytest
+    from vllm.sampling_params import SamplingParams
+
+    from dynamo.vllm.handlers import build_sampling_params
+
+    if not hasattr(SamplingParams(), "routed_experts_prompt_start"):
+        pytest.skip("installed vLLM has no routed_experts_prompt_start support")
+
+    request = {
+        "token_ids": [1, 2, 3],
+        "sampling_options": {},
+        "stop_conditions": {},
+        "output_options": {},
+        "nvext": {"routed_experts_prompt_start": 4},
+    }
+    sp = build_sampling_params(request, default_sampling_params={})
+    assert sp.routed_experts_prompt_start == 4
+
+    # negative / bad values are clamped to 0
+    request["nvext"]["routed_experts_prompt_start"] = -1
+    sp = build_sampling_params(request, default_sampling_params={})
+    assert sp.routed_experts_prompt_start == 0
+
+
 def _make_dynamo_config(**overrides):
     """Build a minimal fake DynamoConfig for update_engine_config_with_dynamo tests."""
     defaults = {

lib/llm/src/preprocessor.rs

@@ -313,6 +313,12 @@ impl OpenAIPreprocessor {
             if nvext.token_data.is_some() {
                 nvext_passthrough.insert("token_in".to_string(), serde_json::Value::Bool(true));
             }
+            if let Some(start) = nvext.routed_experts_prompt_start {
+                nvext_passthrough.insert(
+                    "routed_experts_prompt_start".to_string(),
+                    serde_json::json!(start),
+                );
+            }
         }
 
         if !nvext_passthrough.contains_key("cache_salt")

lib/llm/src/protocols/openai/nvext.rs

@@ -448,6 +448,17 @@ pub struct NvExt {
     #[builder(default, setter(strip_option))]
     pub cache_salt: Option<String>,
 
+    /// RL routed-experts capture offset (MoE expert replay).
+    ///
+    /// When set, Dynamo forwards this to the backend's
+    /// `SamplingParams.routed_experts_prompt_start` so the engine trims the
+    /// leading prompt rows from the returned routing tensor (rather than the
+    /// client trimming after the full-sequence blob crosses the wire). Backends
+    /// without routed-experts capture ignore it.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[builder(default, setter(strip_option))]
+    pub routed_experts_prompt_start: Option<u32>,
+
     /// Extra fields to be included in the response's nvext
     /// This is a list of field names that should be populated in the response
     /// Supported fields include "worker_id", "timing", "routed_experts", "engine_data",