[TRTLLM-16628][refactor] BREAKING: Remove unused LlmRequest fields

cpp/include/tensorrt_llm/batch_manager/llmRequest.h

@@ -112,8 +112,6 @@ class GenericLlmRequest
 
     GenericLlmRequest(RequestIdType requestId, SizeType32 maxNewTokens, std::shared_ptr<VecTokens> const& inputTokens,
         executor::SamplingConfig const& samplingConfig, bool isStreaming,
-        std::optional<TensorPtr> promptEmbeddingTable = std::nullopt,
-        std::optional<SizeType32> promptVocabSize = std::nullopt,
         std::optional<std::shared_ptr<std::vector<std::vector<SizeType32>>>> multimodalHashes = std::nullopt,
         std::optional<std::shared_ptr<std::vector<SizeType32>>> multimodalPositions = std::nullopt,
         std::optional<std::shared_ptr<std::vector<SizeType32>>> multimodalLengths = std::nullopt,
@@ -144,8 +142,6 @@ class GenericLlmRequest
         , mIsStreaming(isStreaming)
         , mOrigPromptLen(mPromptLen)
         , mMaxSentTokenLen(mPromptLen)
-        , mPromptEmbeddingTable(std::move(promptEmbeddingTable))
-        , mPromptVocabSize(promptVocabSize)
         , mMultimodalHashes(std::move(multimodalHashes))
         , mMultimodalPositions(std::move(multimodalPositions))
         , mMultimodalLengths(std::move(multimodalLengths))
@@ -188,8 +184,6 @@ class GenericLlmRequest
 
     GenericLlmRequest(RequestIdType requestId, SizeType32 maxNewTokens, VecTokens const& inputTokens,
         executor::SamplingConfig const& samplingConfig, bool isStreaming,
-        std::optional<TensorPtr> promptEmbeddingTable = std::nullopt,
-        std::optional<SizeType32> promptVocabSize = std::nullopt,
         std::optional<LoraTaskIdType> loraTaskId = std::nullopt, std::optional<TensorPtr> loraWeights = std::nullopt,
         std::optional<TensorPtr> loraConfig = std::nullopt, std::optional<VecTokens> draftTokens = std::nullopt,
         bool excludeInputFromOutput = false, std::optional<VecTokens> encoderInputTokens = std::nullopt,
@@ -203,8 +197,6 @@ class GenericLlmRequest
         , mIsStreaming(isStreaming)
         , mOrigPromptLen(mPromptLen)
         , mMaxSentTokenLen(mPromptLen)
-        , mPromptEmbeddingTable(std::move(promptEmbeddingTable))
-        , mPromptVocabSize(promptVocabSize)
         , mLoraTaskId(loraTaskId)
         , mLoraWeights(std::move(loraWeights))
         , mLoraConfig(std::move(loraConfig))
@@ -271,21 +263,6 @@ class GenericLlmRequest
             }
         }
 
-        auto pTuningConfig = req.getPromptTuningConfig();
-        if (pTuningConfig)
-        {
-            mPromptEmbeddingTable = tensorrt_llm::runtime::ITensor::view(
-                executor::detail::toITensor(pTuningConfig.value().getEmbeddingTable()));
-            TLLM_CHECK(mPromptEmbeddingTable.value()->getShape().nbDims == 2);
-            mPromptVocabSize = mPromptEmbeddingTable.value()->getShape().d[0];
-            mPromptEmbeddingTable.value()->unsqueeze(0);
-
-            if (pTuningConfig->getInputTokenExtraIds())
-            {
-                mInputTokenExtraIds
-                    = std::make_shared<VecTokenExtraIds>(pTuningConfig->getInputTokenExtraIds().value());
-            }
-        }
         auto multimodalInput = req.getMultimodalInput();
         if (multimodalInput)
         {
@@ -776,21 +753,6 @@ class GenericLlmRequest
         mMaxSentTokenLen = maxSentLength;
     }
 
-    [[nodiscard]] std::optional<TensorPtr> getPromptEmbeddingTable() const
-    {
-        return mPromptEmbeddingTable;
-    }
-
-    [[nodiscard]] std::optional<TensorPtr>& getPromptEmbeddingTableMutable()
-    {
-        return mPromptEmbeddingTable;
-    }
-
-    [[nodiscard]] std::optional<SizeType32> getPromptVocabSize() const
-    {
-        return mPromptVocabSize;
-    }
-
     [[nodiscard]] std::optional<std::shared_ptr<std::vector<std::vector<SizeType32>>>> getMultimodalHashes() const
     {
         return mMultimodalHashes;
@@ -846,11 +808,6 @@ class GenericLlmRequest
         mLoraWeights = weights;
     }
 
-    void setPromptVocabSize(SizeType32 size)
-    {
-        mPromptVocabSize = size;
-    }
-
     void clearLoraWeights()
     {
         mLoraWeights = std::nullopt;
@@ -1534,8 +1491,6 @@ class GenericLlmRequest
 
     SizeType32 mMaxSentTokenLen;
 
-    std::optional<TensorPtr> mPromptEmbeddingTable{std::nullopt};
-    std::optional<SizeType32> mPromptVocabSize{std::nullopt};
     std::optional<std::shared_ptr<std::vector<std::vector<SizeType32>>>> mMultimodalHashes{std::nullopt};
     std::optional<std::shared_ptr<std::vector<SizeType32>>> mMultimodalPositions{std::nullopt};
     std::optional<std::shared_ptr<std::vector<SizeType32>>> mMultimodalLengths{std::nullopt};
@@ -1686,16 +1641,6 @@ class GenericLlmRequest
             mEncoderUniqueTokens = encoderUniqueTokens;
         }
 
-        if ((mPromptEmbeddingTable.has_value() && !mPromptVocabSize.has_value())
-            || (!mPromptEmbeddingTable.has_value() && mPromptVocabSize.has_value()))
-        {
-            std::string errStr
-                = "Prompt embedding table and prompt vocab size tensors must both be provided for requests with "
-                  "prompt "
-                  "tuning enabled.";
-            TLLM_THROW(errStr);
-        }
-
         if (!isChild())
         {
             // Initialize result states unless it is a child and a child request should share parent's one.
@@ -1742,8 +1687,6 @@ class LlmRequest : public GenericLlmRequest<runtime::ITensor::SharedPtr>
 
     LlmRequest(RequestIdType requestId, SizeType32 maxNewTokens, std::vector<TokenIdType> inputTokens,
         executor::SamplingConfig const& samplingConfig, bool isStreaming,
-        std::optional<TensorPtr> promptEmbeddingTable = std::nullopt,
-        std::optional<SizeType32> promptVocabSize = std::nullopt,
         std::optional<std::vector<std::vector<SizeType32>>> multimodalHashes = std::nullopt,
         std::optional<std::vector<SizeType32>> multimodalPositions = std::nullopt,
         std::optional<std::vector<SizeType32>> multimodalLengths = std::nullopt,
@@ -1766,7 +1709,7 @@ class LlmRequest : public GenericLlmRequest<runtime::ITensor::SharedPtr>
         std::optional<std::vector<SizeType32>> multimodalRunLengths = std::nullopt,
         std::optional<std::string> cacheSalt = std::nullopt)
         : Base(requestId, maxNewTokens, std::make_shared<std::vector<TokenIdType>>(std::move(inputTokens)),
-            samplingConfig, isStreaming, std::move(promptEmbeddingTable), promptVocabSize,
+            samplingConfig, isStreaming,
             multimodalHashes.has_value()
                 ? std::make_shared<std::vector<std::vector<SizeType32>>>(std::move(multimodalHashes.value()))
                 : std::optional<std::shared_ptr<std::vector<std::vector<SizeType32>>>>(std::nullopt),

cpp/include/tensorrt_llm/executor/executor.h

@@ -252,27 +252,6 @@ class OutputConfig
     std::optional<std::vector<AdditionalModelOutput>> additionalModelOutputs;
 };
 
-/// @brief Configuration for prompt tuning
-class PromptTuningConfig
-{
-public:
-    explicit PromptTuningConfig(
-        Tensor embeddingTable, std::optional<VecTokenExtraIds> inputTokenExtraIds = std::nullopt);
-
-    [[nodiscard]] Tensor getEmbeddingTable() const;
-
-    [[nodiscard]] std::optional<VecTokenExtraIds> getInputTokenExtraIds() const;
-
-private:
-    friend class Serialization;
-    /// @brief The prompt embedding table. Expected shape: [task vocab_size, hidden_size]. Data type must match model
-    /// weights.
-    Tensor mEmbeddingTable;
-
-    /// @brief The input token extra ids for KV Cache reuse when p-tuning is enabled
-    std::optional<VecTokenExtraIds> mInputTokenExtraIds;
-};
-
 /// @brief Multimodal input data class
 class MultimodalInput
 {
@@ -610,7 +589,6 @@ class Request
     /// @param badWords A list of bad words tokens. Each "word" can be composed of multiple tokens
     /// @param stopWords A list of stop words tokens. Each "word" can be composed of multiple tokens
     /// @param embeddingBias The embedding bias tensor. Expected shape is [vocab_size]
-    /// @param pTuningConfig The prompt tuning configuration
     /// @param multimodalInput The multimodal input {multimodalHashes, multimodalPositions, multimodalLengths, optional
     /// exact prompt runs}
     /// @param mRopeConfig The mrope configuration
@@ -641,7 +619,6 @@ class Request
         std::optional<std::list<VecTokens>> badWords = std::nullopt,
         std::optional<std::list<VecTokens>> stopWords = std::nullopt,
         std::optional<Tensor> embeddingBias = std::nullopt,
-        std::optional<PromptTuningConfig> pTuningConfig = std::nullopt,
         std::optional<MultimodalInput> multimodalInput = std::nullopt,
         std::optional<MropeConfig> mRopeConfig = std::nullopt, std::optional<LoraConfig> loraConfig = std::nullopt,
         std::optional<KvCacheRetentionConfig> kvCacheRetentionConfig = std::nullopt,
@@ -672,7 +649,6 @@ class Request
     [[nodiscard]] std::optional<std::list<VecTokens>> getBadWords() const;
     [[nodiscard]] std::optional<std::list<VecTokens>> getStopWords() const;
     [[nodiscard]] std::optional<Tensor> getEmbeddingBias() const;
-    [[nodiscard]] std::optional<PromptTuningConfig> getPromptTuningConfig() const;
     [[nodiscard]] std::optional<MultimodalInput> getMultimodalInput() const;
     [[nodiscard]] std::optional<MropeConfig> getMropeConfig() const;
     [[nodiscard]] std::optional<LoraConfig> getLoraConfig() const;
@@ -699,7 +675,6 @@ class Request
     void setBadWords(std::list<VecTokens> const& badWords);
     void setStopWords(std::list<VecTokens> const& stopWords);
     void setEmbeddingBias(Tensor const& embeddingBias);
-    void setPromptTuningConfig(PromptTuningConfig const& pTuningConfig);
     void setMultimodalInput(MultimodalInput const& multimodalInput);
     void setMropeConfig(MropeConfig const& mRopeConfig);
     void setLoraConfig(LoraConfig const& loraConfig);
@@ -732,12 +707,6 @@ struct Result
     /// @brief The output tokens for each beam
     BeamTokens outputTokenIds;
 
-    /// @brief The cumulative log probabilities. Size beamSize.
-    std::optional<VecLogProbs> cumLogProbs;
-
-    /// @brief The log probabilities for each generated token. Size [beamSize, outputLen]
-    std::optional<std::vector<VecLogProbs>> logProbs;
-
     /// @brief The reason why the model stopped generating tokens for each beam in this request. Size [beamSize].
     /// Currently only supported when beamSize is 1 and when using in-flight batching.
     std::vector<FinishReason> finishReasons;

cpp/include/tensorrt_llm/executor/serialization.h

@@ -66,11 +66,6 @@ class Serialization
     static void serialize(AdditionalModelOutput const& additionalModelOutput, std::ostream& os);
     [[nodiscard]] static size_t serializedSize(AdditionalModelOutput const& additionalModelOutput);
 
-    // PromptTuningConfig
-    [[nodiscard]] static PromptTuningConfig deserializePromptTuningConfig(std::istream& is);
-    static void serialize(PromptTuningConfig const& config, std::ostream& os);
-    [[nodiscard]] static size_t serializedSize(PromptTuningConfig const& config);
-
     // MultimodalInput
     [[nodiscard]] static MultimodalInput deserializeMultimodalInput(std::istream& is);
     static void serialize(MultimodalInput const& multimodalInput, std::ostream& os);

cpp/tensorrt_llm/batch_manager/llmRequest.cpp

@@ -198,7 +198,7 @@ bool LlmRequest::checkTokenIdRange(SizeType32 vocabSize)
 }
 
 void LlmRequest::validate(
-    SizeType32 maxInputLen, SizeType32 maxSequenceLen, SizeType32 maxDraftLen, bool enableKVCacheReuse)
+    SizeType32 maxInputLen, SizeType32 maxSequenceLen, SizeType32 maxDraftLen, bool /*enableKVCacheReuse*/)
 {
     if (mPromptLen > maxInputLen)
     {
@@ -244,27 +244,17 @@ void LlmRequest::validate(
             mPromptLen, mMaxNewTokens, draftLenPerEngineStep, maxSequenceLen, maxNewTokens);
         mMaxNewTokens = maxNewTokens;
     }
-
-    // validate extra ids when enabling kv cache reuse with prompt table
-    if (enableKVCacheReuse && mPromptEmbeddingTable.has_value() && mPromptVocabSize.has_value())
-    {
-        TLLM_CHECK_WITH_INFO(mInputTokenExtraIds.has_value() && mInputTokenExtraIds.value(),
-            "Input token extra ids must be provided when enabling kv cache reuse with prompt table");
-        TLLM_CHECK_WITH_INFO(mInputTokenExtraIds.value()->size() == static_cast<size_t>(mOrigPromptLen),
-            "inputTokenExtraIds vector size (%lu) must be the same as input token vector size (%lu).",
-            mInputTokenExtraIds.value()->size(), static_cast<size_t>(mOrigPromptLen));
-    }
 }
 
 LlmRequest::LlmRequest(LlmRequest const& parent, RequestIdType requestId)
     : LlmRequest(requestId, parent.mMaxNewTokens, parent.mTokens.at(0), parent.mSamplingConfig, parent.mIsStreaming,
-        parent.mPromptEmbeddingTable, parent.mPromptVocabSize, copyOptionalInput(parent.mMultimodalHashes),
-        copyOptionalInput(parent.mMultimodalPositions), copyOptionalInput(parent.mMultimodalLengths),
-        copyOptionalInput(parent.mMultimodalUuids), parent.mLoraTaskId, parent.mLoraWeights, parent.mLoraConfig,
-        parent.mKvCacheRetentionConfig, parent.mDraftTokens ? *parent.mDraftTokens : VecTokens{},
-        parent.mExcludeInputFromOutput, copyOptionalInput(parent.mEncoderTokens), parent.mPriority,
-        parent.mEncoderInputLength, parent.mEncoderOutputLength, parent.mLlmRequestType,
-        copyOptionalInput(parent.mInputTokenExtraIds), parent.mReturnPerfMetrics, parent.mContextPhaseParams,
+        copyOptionalInput(parent.mMultimodalHashes), copyOptionalInput(parent.mMultimodalPositions),
+        copyOptionalInput(parent.mMultimodalLengths), copyOptionalInput(parent.mMultimodalUuids), parent.mLoraTaskId,
+        parent.mLoraWeights, parent.mLoraConfig, parent.mKvCacheRetentionConfig,
+        parent.mDraftTokens ? *parent.mDraftTokens : VecTokens{}, parent.mExcludeInputFromOutput,
+        copyOptionalInput(parent.mEncoderTokens), parent.mPriority, parent.mEncoderInputLength,
+        parent.mEncoderOutputLength, parent.mLlmRequestType, copyOptionalInput(parent.mInputTokenExtraIds),
+        parent.mReturnPerfMetrics, parent.mContextPhaseParams,
         parent.mReturnPerfMetrics ? std::make_optional(parent.mPerfMetrics.timingMetrics.arrivalTime) : std::nullopt,
         parent.mAgentHierarchy, copyOptionalInput(parent.mMultimodalItemRunCuOffsets),
         copyOptionalInput(parent.mMultimodalRunPositions), copyOptionalInput(parent.mMultimodalRunLengths),

cpp/tensorrt_llm/executor/CMakeLists.txt

@@ -80,7 +80,6 @@ add_library(
   loraConfig.cpp
   mropeConfig.cpp
   multimodalInput.cpp
-  promptTuningConfig.cpp
   request.cpp
   response.cpp
   serialization.cpp

cpp/tensorrt_llm/executor/request.cpp

@@ -28,19 +28,18 @@ Request::Request(VecTokens inputTokenIds, SizeType32 maxTokens, bool streaming,
     OutputConfig const& outputConfig, std::optional<SizeType32> const& endId,
     std::optional<std::vector<SizeType32>> positionIds, std::optional<std::list<VecTokens>> badWords,
     std::optional<std::list<VecTokens>> stopWords, std::optional<Tensor> embeddingBias,
-    std::optional<PromptTuningConfig> pTuningConfig, std::optional<MultimodalInput> multimodalInput,
-    std::optional<MropeConfig> mRopeConfig, std::optional<LoraConfig> loraConfig,
-    std::optional<KvCacheRetentionConfig> kvCacheRetentionConfig, std::optional<VecTokens> encoderInputTokenIds,
-    std::optional<IdType> clientId, bool returnAllGeneratedTokens, float priority, RequestType type,
-    std::optional<ContextPhaseParams> contextPhaseParams, std::optional<Tensor> encoderInputFeatures,
-    std::optional<SizeType32> encoderOutputLength, std::optional<GuidedDecodingParams> guidedDecodingParams,
-    std::optional<MillisecondsType> allottedTimeMs, std::optional<IdType> disaggRequestId,
-    std::optional<std::string> cacheSalt)
+    std::optional<MultimodalInput> multimodalInput, std::optional<MropeConfig> mRopeConfig,
+    std::optional<LoraConfig> loraConfig, std::optional<KvCacheRetentionConfig> kvCacheRetentionConfig,
+    std::optional<VecTokens> encoderInputTokenIds, std::optional<IdType> clientId, bool returnAllGeneratedTokens,
+    float priority, RequestType type, std::optional<ContextPhaseParams> contextPhaseParams,
+    std::optional<Tensor> encoderInputFeatures, std::optional<SizeType32> encoderOutputLength,
+    std::optional<GuidedDecodingParams> guidedDecodingParams, std::optional<MillisecondsType> allottedTimeMs,
+    std::optional<IdType> disaggRequestId, std::optional<std::string> cacheSalt)
     : mImpl(std::make_unique<Impl>(std::move(inputTokenIds), maxTokens, streaming, samplingConfig, outputConfig, endId,
         std::move(positionIds), std::move(badWords), std::move(stopWords), std::move(embeddingBias),
-        std::move(pTuningConfig), std::move(multimodalInput), std::move(mRopeConfig), std::move(loraConfig),
-        std::move(kvCacheRetentionConfig), std::move(encoderInputTokenIds), clientId, returnAllGeneratedTokens,
-        priority, type, std::move(contextPhaseParams), std::move(encoderInputFeatures), encoderOutputLength,
+        std::move(multimodalInput), std::move(mRopeConfig), std::move(loraConfig), std::move(kvCacheRetentionConfig),
+        std::move(encoderInputTokenIds), clientId, returnAllGeneratedTokens, priority, type,
+        std::move(contextPhaseParams), std::move(encoderInputFeatures), encoderOutputLength,
         std::move(guidedDecodingParams), allottedTimeMs, disaggRequestId, std::move(cacheSalt)))
 {
 }
@@ -120,11 +119,6 @@ std::optional<Tensor> Request::getEmbeddingBias() const
     return mImpl->getEmbeddingBias();
 }
 
-std::optional<PromptTuningConfig> Request::getPromptTuningConfig() const
-{
-    return mImpl->getPromptTuningConfig();
-}
-
 std::optional<MultimodalInput> Request::getMultimodalInput() const
 {
     return mImpl->getMultimodalInput();
@@ -245,11 +239,6 @@ void Request::setEmbeddingBias(Tensor const& embeddingBias)
     mImpl->setEmbeddingBias(embeddingBias);
 }
 
-void Request::setPromptTuningConfig(PromptTuningConfig const& pTuningConfig)
-{
-    mImpl->setPromptTuningConfig(pTuningConfig);
-}
-
 void Request::setMultimodalInput(MultimodalInput const& multimodalInput)
 {
     mImpl->setMultimodalInput(multimodalInput);