[None][fix] Fix early_stopping type and plumb through Triton ensemble…

triton_backend/all_models/disaggregated_serving/disaggregated_serving_bls/config.pbtxt

@@ -171,7 +171,7 @@ input [
   },
   {
     name: "early_stopping"
-    data_type: TYPE_BOOL
+    data_type: TYPE_INT32
     dims: [ 1 ]
     optional: true
   },

triton_backend/all_models/gpt/ensemble/config.pbtxt

@@ -58,6 +58,12 @@ input [
     dims: [ 1 ]
     optional: true
   },
+  {
+    name: "early_stopping"
+    data_type: TYPE_INT32
+    dims: [ 1 ]
+    optional: true
+  },
   {
     name: "prompt_ignore_length"
     data_type: TYPE_INT32
@@ -187,6 +193,10 @@ ensemble_scheduling {
           key: "len_penalty"
           value: "length_penalty"
       }
+      input_map {
+          key: "early_stopping"
+          value: "early_stopping"
+      }
       input_map {
           key: "prompt_ignore_length"
           value: "prompt_ignore_length"

triton_backend/all_models/gpt/tensorrt_llm/config.pbtxt

@@ -71,6 +71,13 @@ input [
     reshape: { shape: [ ] }
     optional: true
   },
+  {
+    name: "early_stopping"
+    data_type: TYPE_INT32
+    dims: [ 1 ]
+    reshape: { shape: [ ] }
+    optional: true
+  },
   {
     name: "prompt_ignore_length"
     data_type: TYPE_INT32

triton_backend/all_models/inflight_batcher_llm/ensemble/config.pbtxt

@@ -104,6 +104,12 @@ input [
     dims: [ 1 ]
     optional: true
   },
+  {
+    name: "early_stopping"
+    data_type: TYPE_INT32
+    dims: [ 1 ]
+    optional: true
+  },
   {
     name: "prompt_ignore_length"
     data_type: TYPE_INT32
@@ -521,6 +527,10 @@ ensemble_scheduling {
           key: "len_penalty"
           value: "length_penalty"
       }
+      input_map {
+          key: "early_stopping"
+          value: "early_stopping"
+      }
       input_map {
           key: "prompt_ignore_length"
           value: "prompt_ignore_length"

triton_backend/all_models/inflight_batcher_llm/tensorrt_llm/config.pbtxt

@@ -214,7 +214,7 @@ input [
   },
   {
     name: "early_stopping"
-    data_type: TYPE_BOOL
+    data_type: TYPE_INT32
     dims: [ 1 ]
     reshape: { shape: [ ] }
     optional: true

triton_backend/all_models/inflight_batcher_llm/tensorrt_llm_bls/1/lib/decode.py

@@ -76,6 +76,7 @@ class Request:
     top_p: Optional[np.ndarray] = None
     temperature: Optional[np.ndarray] = None
     length_penalty: Optional[np.ndarray] = None
+    early_stopping: Optional[np.ndarray] = None
     prompt_ignore_length: Optional[np.ndarray] = None
     repetition_penalty: Optional[np.ndarray] = None
     min_tokens: Optional[np.ndarray] = None

triton_backend/all_models/inflight_batcher_llm/tensorrt_llm_bls/1/lib/triton_decoder.py

@@ -90,10 +90,11 @@ def __init__(self,
             "image_bytes_input", "image_url_input", "video_bytes_input",
             "max_tokens", "bad_words", "stop_words", "end_id", "pad_id",
             "top_k", "top_p", "temperature", "length_penalty",
-            "prompt_ignore_length", "repetition_penalty", "min_tokens",
-            "presence_penalty", "frequency_penalty", "seed", "return_log_probs",
-            "return_context_logits", "return_generation_logits", "beam_width",
-            "stream", "prompt_embedding_table", "prompt_vocab_size",
+            "prompt_ignore_length", "early_stopping", "repetition_penalty",
+            "min_tokens", "presence_penalty", "frequency_penalty", "seed",
+            "return_log_probs", "return_context_logits",
+            "return_generation_logits", "beam_width", "stream",
+            "prompt_embedding_table", "prompt_vocab_size",
             "prompt_table_extra_id", "embedding_bias_words",
             "embedding_bias_weights", "num_draft_tokens", "use_draft_logits",
             "lora_task_id", "lora_weights", "lora_config",
@@ -104,13 +105,14 @@ def __init__(self,
 
         self.__undo_reshape_whitelist = {
             "max_tokens", "end_id", "pad_id", "top_k", "top_p", "temperature",
-            "length_penalty", "prompt_ignore_length", "repetition_penalty",
-            "min_tokens", "presence_penalty", "frequency_penalty", "seed",
-            "return_log_probs", "return_context_logits",
-            "return_generation_logits", "beam_width", "stream",
-            "prompt_vocab_size", "num_draft_tokens", "use_draft_logits",
-            "exclude_input_in_output", "return_perf_metrics", "lora_weights",
-            "lora_config", "lora_task_id", "return_num_input_tokens",
+            "length_penalty", "prompt_ignore_length", "early_stopping",
+            "repetition_penalty", "min_tokens", "presence_penalty",
+            "frequency_penalty", "seed", "return_log_probs",
+            "return_context_logits", "return_generation_logits", "beam_width",
+            "stream", "prompt_vocab_size", "num_draft_tokens",
+            "use_draft_logits", "exclude_input_in_output",
+            "return_perf_metrics", "lora_weights", "lora_config",
+            "lora_task_id", "return_num_input_tokens",
             "return_num_output_tokens"
         }
 
@@ -456,6 +458,7 @@ def _get_llm_tensors_from_request(
             "temperature": "temperature",
             "length_penalty": "len_penalty",
             "prompt_ignore_length": "prompt_ignore_length",
+            "early_stopping": "early_stopping",
             "repetition_penalty": "repetition_penalty",
             "min_tokens": "min_tokens",
             "presence_penalty": "presence_penalty",

triton_backend/all_models/inflight_batcher_llm/tensorrt_llm_bls/config.pbtxt

@@ -134,6 +134,12 @@ input [
     dims: [ 1 ]
     optional: true
   },
+  {
+    name: "early_stopping"
+    data_type: TYPE_INT32
+    dims: [ 1 ]
+    optional: true
+  },
   {
     name: "prompt_ignore_length"
     data_type: TYPE_INT32

triton_backend/all_models/multimodal/ensemble/config.pbtxt

@@ -123,6 +123,12 @@ input [
     dims: [ 1 ]
     optional: true
   },
+  {
+    name: "early_stopping"
+    data_type: TYPE_INT32
+    dims: [ 1 ]
+    optional: true
+  },
   {
     name: "prompt_ignore_length"
     data_type: TYPE_INT32
@@ -518,6 +524,10 @@ ensemble_scheduling {
           key: "len_penalty"
           value: "length_penalty"
       }
+      input_map {
+          key: "early_stopping"
+          value: "early_stopping"
+      }
       input_map {
           key: "prompt_ignore_length"
           value: "prompt_ignore_length"

triton_backend/inflight_batcher_llm/tests/utilsTest.cpp

@@ -382,7 +382,11 @@ std::optional<tensorrt_llm::executor::Request> getRequest(
     pushTensor<int32_t>(inputsTensors, InputFieldsNames::topPResetIds, nvinfer1::DataType::kINT32, {1}, {25});
     pushTensor<float>(inputsTensors, InputFieldsNames::temperature, nvinfer1::DataType::kFLOAT, {1}, {0.3});
     pushTensor<float>(inputsTensors, InputFieldsNames::lengthPenalty, nvinfer1::DataType::kFLOAT, {1}, {0.4});
-    pushTensor<int32_t>(inputsTensors, InputFieldsNames::earlyStopping, nvinfer1::DataType::kINT32, {1}, {4});
+    // early_stopping is a HuggingFace-style tri-state: 0=heuristic, 1=fast,
+    // 2=stop only when all beams emit <eos>. Push 2 to lock in INT32 wire-format
+    // for the canonical-beam-search mode (was silently truncated under the old
+    // TYPE_BOOL config declaration; cf. PR fixing the wire-protocol type).
+    pushTensor<int32_t>(inputsTensors, InputFieldsNames::earlyStopping, nvinfer1::DataType::kINT32, {1}, {2});
     pushTensor<float>(inputsTensors, InputFieldsNames::repetitionPenalty, nvinfer1::DataType::kFLOAT, {1}, {0.8});
     pushTensor<int32_t>(inputsTensors, InputFieldsNames::minTokens, nvinfer1::DataType::kINT32, {1}, {45});
     pushTensor<float>(inputsTensors, InputFieldsNames::beamSearchDiversityRate, nvinfer1::DataType::kFLOAT, {1}, {0.1});
@@ -580,7 +584,7 @@ void checkRequest(tensorrt_llm::executor::Request const& request,
     EXPECT_EQ(samplingConfig.getTopPResetIds().value(), 25);
     EXPECT_EQ(samplingConfig.getTemperature().value(), 0.3f);
     EXPECT_EQ(samplingConfig.getLengthPenalty().value(), 0.4f);
-    EXPECT_EQ(samplingConfig.getEarlyStopping().value(), 4);
+    EXPECT_EQ(samplingConfig.getEarlyStopping().value(), 2);
     EXPECT_EQ(samplingConfig.getRepetitionPenalty().value(), 0.8f);
     EXPECT_EQ(samplingConfig.getMinTokens().value(), 45);
     EXPECT_EQ(samplingConfig.getBeamSearchDiversityRate().value(), 0.1f);