/* * SPDX-FileCopyrightText: Copyright (c) 1993-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved. * SPDX-License-Identifier: Apache-2.0 * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "checkMacrosPlugin.h" #include "tensorrt_llm/common/cublasMMWrapper.h" #include "tensorrt_llm/common/logger.h" #include "tensorrt_llm/common/quantization.h" #include "tensorrt_llm/common/stringUtils.h" #include "tensorrt_llm/kernels/contextFusedMultiHeadAttention/fmhaRunner.h" #include "tensorrt_llm/kernels/contextFusedMultiHeadAttention/fused_multihead_attention_common.h" #include "tensorrt_llm/kernels/gptKernels.h" #include "tensorrt_llm/plugins/common/plugin.h" #include "tensorrt_llm/plugins/gptAttentionCommon/gptAttentionCommon.h" #include #include #include #include #include #include namespace tensorrt_llm::plugins { // batch_size = num_ctx_requests + num_gen_requests * beam_width // num_ctx_requests = number of context requests (single sequence per request). // num_gen_requests = number of generation requests (beam_width sequences per request). // Context sequences have to appear first, generation sequences after // inputs (see GPTAttentionPlugin::isEntryUsed for when each tensor is actually used) // 0. input_tensor [batch_size, seq_len, local_hidden_size + 2 * local_num_kv_heads * head_size] or // [num_tokens, local_hidden_size + 2 * local_num_kv_heads * head_size] when // enable_remove_input_padding // 1. sequence_length [batch_size] (optional) // 2. host_past_key_value_lengths [batch_size] (int32) (optional) // 3. host_max_attention_window_sizes [num_layers] (int32) // 4. host_sink_token_length [1] (int32) // 5. context_lengths [batch_size] // 6. cache_indir [num_gen_requests, beam_width, memory_max_len] (required in beamsearch) (optional) // 7. host_request_types [batch_size] int32. 0: context; 1: generation: 2: none. When not in inflight-batching // mode, // all elements must be identical. // 8. past_key_value_pool [batch_size, 2, local_num_kv_heads, max_seq_len, head_size] or // block_offsets [batch_size, 2, max_blocks_per_seq] if paged kv cache (optional) // 8.1 host_block_offsets [batch_size, 2, max_blocks_per_seq] if paged kv cache (optional) // 8.2 host_pool_pointers [2] if paged kv cache (optional) // 9. kv_cache_quantization_scale [1] (optional) // 10. kv_cache_dequantization_scale [1] (optional) // 11. attention_output_quantization_scale [1] (on device, optional) // 12. attention_mask [num_tokens, kv_seqlen] (on device, bool, optional) // 13. attention_packed_mask [num_tokens, kv_seqlen / 32] (on device, uint32_t, optional) // - pack masks by encoding multiple mask positions into a single 32-bit unsigned integer. // - see kernels/contextMultiHeadAttention/fmhaPackedMask.cpp for more details. // 14. rotary_inv_freq [head_size / 2] or [head_size] (longrope type) (float) (on device, optional) // 15. rotary_cos_sin [max_num_embedding_positions, 2] (float) (on device, optional) // 16. alibi_slopes [num_heads] (optional for ALiBi position embedding) // 17. relative_attention_bias [num_heads] (optional for ALiBi position embedding) // 18. host_context_lengths [batch_size] int32. (optional, required when remove_input_padding is true) // 19. qkv_bias (optional) [local_hidden_size * 3] // 20. spec_decoding_generation_lengths (optional, required when medusa is enabled) (int32_t) [batch_size] // 21. spec_decoding_packed_mask (optional, required when medusa is enabled) (int32_t) [num_tokens, packed_mask_dim] // packed_mask_dim = divUp(max_num_spec_decoding_tokens + 1, 32) // 22. spec_decoding_position_offsets (optional, required when medusa is enabled) (int32_t) [batch_size, // max_num_spec_decoding_tokens + 1] // 23. spec_decoding_use (optional, bool) [1]: If it is set as true, enable speculative decoding // 24. long_rope_rotary_inv_freq [head / 2] (float) (on device, optional) // 25. long_rope_rotary_cos_sin [max_num_embedding_positions, 2] (float) (on device, optional) // 26. host_runtime_perf_knobs (int64) // 27. host_context_progress (void*) // 28. position_id_tensor(MLA) [total_tokens], used for rope embedding in MLA // 29. q_a_proj_tensor(MLA) [hidden_dim, c_q_dim + c_k_dim + ropd_dim], used to proj compacted QKV // 30. q_a_layernorm_tensor(MLA) [c_q_dim], rmsnorm weight for compacted q // 31. q_b_proj_tensor(MLA) [c_q_dim, head_num * head_size], weight for companted q to q in context // 32. kv_a_proj_with_mqa_tensor(MLA) [c_q_dim, head_num * (c_k_dim + rope_dim)], weight for companted q to kdim in // generation // 33. kv_a_layernorm_tensor(MLA) [c_k_dim], rmsnorm weight for compacted kv // 34. kv_b_proj_tensor(MLA) [c_k_dim, head_num * 2 * (head_size - rope_dim)], weight for compacted kv to kv in // context // 35. skip_attn (optional, bool) [1]: If it is set as true, skip the atteniton plugin and return // directly. // // outputs // output_tensor [batch_size, seq_len, local_hidden_size] // present_key_value_pool (optional if not paged kv cache) [batch_size, 2, local_num_kv_heads, max_seq_len, // head_size] class GPTAttentionPlugin : public GPTAttentionPluginCommon { public: GPTAttentionPlugin(int layer_idx, int num_heads, int vision_start, int vision_length, int num_kv_heads, int num_kv_heads_origin, int head_size, int unidirectional, float q_scaling, float attn_logit_softcapping_scale, tensorrt_llm::kernels::PositionEmbeddingType position_embedding_type, int rotary_embedding_dim, // for RoPE. 0 for non-RoPE float rotary_embedding_base, tensorrt_llm::kernels::RotaryScalingType rotary_embedding_scale_type, float rotary_embedding_scale, float rotary_embedding_short_m_scale, float rotary_embedding_long_m_scale, int rotary_embedding_max_positions, int rotary_embedding_original_max_positions, int tp_size, int tp_rank, // for ALiBi bool unfuse_qkv_gemm, // for AutoPP bool use_logn_scaling, // for LognScaling tensorrt_llm::kernels::ContextFMHAType context_fmha_type, int kv_cache_quant_mode, bool remove_input_padding, tensorrt_llm::kernels::AttentionMaskType mask_type, tensorrt_llm::kernels::BlockSparseParams block_sparse_params, bool paged_kv_cache, int tokens_per_block, nvinfer1::DataType type, int32_t max_context_length, bool qkv_bias_enabled, bool cross_attention = false, int max_distance = 0, bool pos_shift_enabled = false, bool dense_context_fmha = false, bool use_paged_context_fmha = true, bool use_fp8_context_fmha = true, bool has_full_attention_mask = false, bool use_cache = true, bool is_spec_decoding_enabled = false, bool spec_decoding_is_generation_length_variable = false, int spec_decoding_max_generation_length = 1, bool is_mla_enabled = false, int q_lora_rank = 0, int kv_lora_rank = 0, int qk_nope_head_dim = 0, int qk_rope_head_dim = 0, int v_head_dim = 0, bool fuse_fp4_quant = false, bool skip_attn = false, int cp_size = 1, int cp_rank = 0, std::set cp_group = {}); GPTAttentionPlugin(void const* data, size_t length); ~GPTAttentionPlugin() override = default; // IPluginV2DynamicExt Methods nvinfer1::DimsExprs getOutputDimensions(int outputIndex, nvinfer1::DimsExprs const* inputs, int nbInputs, nvinfer1::IExprBuilder& exprBuilder) noexcept override; bool supportsFormatCombination( int pos, nvinfer1::PluginTensorDesc const* inOut, int nbInputs, int nbOutputs) noexcept override; size_t getWorkspaceSize(nvinfer1::PluginTensorDesc const* inputs, int nbInputs, nvinfer1::PluginTensorDesc const* outputs, int nbOutputs) const noexcept override; int enqueue(nvinfer1::PluginTensorDesc const* inputDesc, nvinfer1::PluginTensorDesc const* outputDesc, void const* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) noexcept override; template int enqueueImpl(nvinfer1::PluginTensorDesc const* inputDesc, nvinfer1::PluginTensorDesc const* outputDesc, void const* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream); template int enqueueDispatchKVCacheType(nvinfer1::PluginTensorDesc const* inputDesc, nvinfer1::PluginTensorDesc const* outputDesc, void const* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream); template void configurePluginImpl(nvinfer1::DynamicPluginTensorDesc const* in, int nbInputs, nvinfer1::DynamicPluginTensorDesc const* out, int nbOutputs) noexcept; template void configurePluginDispatchKVCacheType(nvinfer1::DynamicPluginTensorDesc const* in, int nbInputs, nvinfer1::DynamicPluginTensorDesc const* out, int nbOutputs) noexcept; void configurePlugin(nvinfer1::DynamicPluginTensorDesc const* in, int nbInputs, nvinfer1::DynamicPluginTensorDesc const* out, int nbOutputs) noexcept override; // IPluginV2Ext Methods nvinfer1::DataType getOutputDataType( int index, nvinfer1::DataType const* inputTypes, int nbInputs) const noexcept override; // IPluginV2 Methods char const* getPluginType() const noexcept override; char const* getPluginVersion() const noexcept override; int getNbOutputs() const noexcept override; //! This is called on every trt ExecutionContext creation by TRT //! Note TRT does not call the initialize on cloned plugin, so clone internally should do initialization. GPTAttentionPlugin* clone() const noexcept override; size_t getSerializationSize() const noexcept override; void serialize(void* buffer) const noexcept override; private: template int enqueueSome(int32_t seqIdxBeg, int32_t localNbSeq, int32_t tokenIdxBeg, int32_t localNbTokens, nvinfer1::PluginTensorDesc const* inputDesc, nvinfer1::PluginTensorDesc const* outputDesc, void const* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream); using IndexType = std::int32_t; std::vector mEntryIdx; enum class IdxEntry : size_t { QKV_TENSOR, K_TENSOR, V_TENSOR, ATTENTION_MASK, ATTENTION_PACKED_MASK, SEQUENCE_LENGTH, HOST_PAST_KEY_VALUE_LENGTHS, HOST_MAX_ATTENTION_WINDOW, HOST_SINK_TOKEN_LENGTH, CONTEXT_LENGTHS, CACHE_INDIR, REQUEST_TYPES, KV_CACHE_BLOCK_OFFSETS, HOST_KV_CACHE_BLOCK_OFFSETS, HOST_KV_CACHE_POOL_POINTERS, HOST_KV_CACHE_POOL_MAPPING, PAST_KEY_VALUE, KV_CACHE_QUANTIZATION_SCALE, KV_CACHE_DEQUANTIZATION_SCALE, ATTENTION_OUTPUT_QUANTIZATION_SCALE, ATTENTION_OUTPUT_SF_SCALE, ROTARY_INV_FREQ, ROTARY_COS_SIN, ALIBI_SLOPES, RELATIVE_ATTENTION_BIAS, CROSS_KV, CROSS_KV_LENGTH, ENCODER_INPUT_LENGTH, HOST_CONTEXT_LENGTH, QKV_BIAS_TENSOR, SPEC_DECODING_GENERATION_LENGTHS, SPEC_DECODING_PACKED_MASK, SPEC_DECODING_POSITION_OFFSETS, SPEC_DECODING_USE, LONG_ROPE_ROTARY_INV_FREQ, LONG_ROPE_ROTARY_COS_SIN, MROPE_ROTARY_COS_SIN, MROPE_POSITION_DELTAS, HOST_RUNTIME_PERF_KNOBS, HOST_CONTEXT_PROGRESS, MLA_Q_B_PROJ_TENSOR, MLA_KV_B_PROJ_TENSOR, MLA_K_B_PROJ_TRANS_TENSOR, SKIP_ATTN, LOGN_SCALING, ENUM_SIZE, // Used to count the number of IdxEntry, must put in last }; std::string toString(IdxEntry const& entry) const; bool isEntryUsed(IdxEntry const& entry) const; void initEntryIdx(); IndexType getIdx(IdxEntry const& entry) const; // Get generation input sequence length (might be larger than 1 in the speculative decoding mode). int getGenerationInputSequenceLength( nvinfer1::PluginTensorDesc const* inputDesc, int32_t localNbSeq, int32_t localNbTokens) const; }; class GPTAttentionPluginCreator : public GPTAttentionPluginCreatorCommon { public: GPTAttentionPluginCreator(); char const* getPluginName() const noexcept override; char const* getPluginVersion() const noexcept override; nvinfer1::PluginFieldCollection const* getFieldNames() noexcept override; nvinfer1::IPluginV2* createPlugin(char const* name, nvinfer1::PluginFieldCollection const* fc) noexcept override; nvinfer1::IPluginV2* deserializePlugin( char const* name, void const* serialData, size_t serialLength) noexcept override; }; } // namespace tensorrt_llm::plugins