Send mm_payload with return_mm_kwargs on vLLM render copies

pkg/epp/framework/plugins/requestcontrol/dataproducer/tokenizer/README.md

@@ -105,7 +105,7 @@ tools, or reconstruct content from routing projections. Model rewrites happen
 before token production and apply to both rendering and forwarding.
 `vllm.prefillOnly` rewrites only the output budget fields of the render copy;
 see [Render-only output budget](#render-only-output-budget). `vllm.omitMMKwargs`
-adds one field to the render copy; see
+adds two fields to the render copy; see
 [Render responses without multimodal tensors](#render-responses-without-multimodal-tensors).
 
 The parsed payload keeps nested objects, arrays, Completions `prompt`, and
@@ -497,11 +497,13 @@ The producer reads token IDs, `mm_hashes`, and `mm_placeholders` from a render
 response. By default vLLM also returns the processed multimodal tensors in
 `features.kwargs_data`, which the producer discards and which can reach tens of
 megabytes per request. Set `vllm.omitMMKwargs: true` to add
-`"return_mm_kwargs": false` to the render copy, so vLLM leaves the tensors out
-and skips serializing them. The inference payload is unchanged. The option
-requires a vLLM version with the `return_mm_kwargs` render field; other
-versions ignore the field and return the tensors. It is false by default and
-works with `vllm.url` and `vllm.endpointDiscovery`.
+`"return_mm_kwargs": false` and `"mm_payload": "none"` to the render copy, so
+vLLM leaves the tensors out and skips serializing them. The inference payload
+is unchanged. vLLM v0.32 reads `return_mm_kwargs`; versions with
+vllm-project/vllm#61427 read `mm_payload`, which takes precedence. Values the
+client sent for either field are overridden. vLLM versions without these fields
+ignore them and return the tensors. The option is false by default and works
+with `vllm.url` and `vllm.endpointDiscovery`.
 
 A complete sample config that pairs this with `precise-prefix-cache-producer` and `prefix-cache-scorer` is at [`deploy/config/sim-epp-tokenizer-vllm-http-config.yaml`](../../../../../../../deploy/config/sim-epp-tokenizer-vllm-http-config.yaml).
 

pkg/epp/framework/plugins/requestcontrol/dataproducer/tokenizer/model_limits_test.go

@@ -342,10 +342,10 @@ func TestModelLimitEligibility(t *testing.T) {
 
 func TestOmitMMKwargs(t *testing.T) {
 	for _, tc := range []struct{ name, payload, want string }{
-		{"appends", `{"model":"m","messages":[]}`, `{"model":"m","messages":[],"return_mm_kwargs":false}`},
-		{"empty object", `{ }`, `{ "return_mm_kwargs":false}`},
-		{"trailing space", "{\"model\":\"m\"}\n", "{\"model\":\"m\",\"return_mm_kwargs\":false}\n"},
-		{"client value loses", `{"return_mm_kwargs":true}`, `{"return_mm_kwargs":true,"return_mm_kwargs":false}`},
+		{"appends", `{"model":"m","messages":[]}`, `{"model":"m","messages":[],"return_mm_kwargs":false,"mm_payload":"none"}`},
+		{"empty object", `{ }`, `{ "return_mm_kwargs":false,"mm_payload":"none"}`},
+		{"trailing space", "{\"model\":\"m\"}\n", "{\"model\":\"m\",\"return_mm_kwargs\":false,\"mm_payload\":\"none\"}\n"},
+		{"client values lose", `{"return_mm_kwargs":true,"mm_payload":"tensors"}`, `{"return_mm_kwargs":true,"mm_payload":"tensors","return_mm_kwargs":false,"mm_payload":"none"}`},
 	} {
 		t.Run(tc.name, func(t *testing.T) {
 			out, err := omitMMKwargs([]byte(tc.payload))
@@ -372,6 +372,7 @@ func TestOmitMMKwargsAppliesToRawPayload(t *testing.T) {
 	_, _, err = renderer.RenderChat(context.Background(), raw)
 	require.NoError(t, err)
 	assert.Equal(t, "false", string(seen["return_mm_kwargs"]))
+	assert.Equal(t, `"none"`, string(seen["mm_payload"]))
 	assert.Equal(t, "1", string(seen["max_tokens"]))
 	assert.Equal(t, `[{"role":"user","content":"hi"}]`, string(seen["messages"]))
 }

pkg/epp/framework/plugins/requestcontrol/dataproducer/tokenizer/vllm_http.go

@@ -437,23 +437,26 @@ func renderOnlyBudget(payload []byte) ([]byte, error) {
 	return json.Marshal(envelope)
 }
 
-var omitMMKwargsField = []byte(`"return_mm_kwargs":false`)
+// omitMMKwargsFields carries both spellings of the vLLM render option. vLLM v0.32 reads
+// return_mm_kwargs; versions with vllm-project/vllm#61427 read mm_payload, which takes precedence
+// over return_mm_kwargs. A version without a field ignores it.
+var omitMMKwargsFields = []byte(`"return_mm_kwargs":false,"mm_payload":"none"`)
 
-// omitMMKwargs adds "return_mm_kwargs":false before the closing brace of the render copy of a JSON
+// omitMMKwargs adds omitMMKwargsFields before the closing brace of the render copy of a JSON
 // object. It splices instead of decoding, because the body can carry megabytes of base64 images.
-// A later duplicate key wins in vLLM's parser, so a client-sent value cannot override it.
+// A later duplicate key wins in vLLM's parser, so client-sent values cannot override them.
 func omitMMKwargs(payload []byte) ([]byte, error) {
 	end := bytes.LastIndexByte(payload, '}')
 	head := bytes.TrimSpace(payload[:max(end, 0)])
 	if end < 0 || len(head) == 0 || head[0] != '{' || len(bytes.TrimSpace(payload[end+1:])) != 0 {
 		return nil, errors.New("render payload is not a JSON object")
 	}
-	out := make([]byte, 0, len(payload)+len(omitMMKwargsField)+1)
+	out := make([]byte, 0, len(payload)+len(omitMMKwargsFields)+1)
 	out = append(out, payload[:end]...)
 	if head[len(head)-1] != '{' {
 		out = append(out, ',')
 	}
-	out = append(out, omitMMKwargsField...)
+	out = append(out, omitMMKwargsFields...)
 	return append(out, payload[end:]...), nil
 }