Merge branch 'main' into adithyare/unfused_lora

Dockerfile

@@ -67,6 +67,7 @@ WORKDIR /workspace/
 RUN git clone https://github.com/NVIDIA/Megatron-LM.git && \
   cd Megatron-LM && \
   git checkout 36e9b6bf3d8034b10c9bbd9fc357c2df2bd1515c && \
+  git cherry-pick -n e69187bc3679ea5841030a165d587bb48b56ee77 && \
   pip install .
 
 # Performance optimizations for distributed optimizer: https://github.com/NVIDIA/apex/pull/1771
@@ -133,7 +134,7 @@ RUN pip install flash-attn
 # install numba for latest containers
 RUN pip install numba>=0.57.1
 # install ammo
-RUN pip install nvidia-ammo~=0.7.0 --extra-index-url https://pypi.nvidia.com --no-cache-dir
+RUN pip install nvidia-ammo~=0.9.0 --extra-index-url https://pypi.nvidia.com --no-cache-dir
 
 # copy nemo source into a scratch image
 FROM scratch as nemo-src

Jenkinsfile

@@ -97,7 +97,7 @@ pipeline {
 
     stage('AMMO installation') {
       steps {
-         sh 'pip install nvidia-ammo~=0.7.0 --extra-index-url https://pypi.nvidia.com --no-cache-dir'
+         sh 'pip install nvidia-ammo~=0.9.0 --extra-index-url https://pypi.nvidia.com --no-cache-dir'
       }
     }
 

docs/source/index.rst

@@ -60,6 +60,7 @@ For more information, browse the developer docs for your area of interest in the
    nlp/models
    nlp/machine_translation/machine_translation
    nlp/megatron_onnx_export
+   nlp/quantization
    nlp/api
 
 

docs/source/nlp/quantization.rst

@@ -3,35 +3,35 @@
 Quantization
 ==========================
 
-Post Training Quantization (PTQ)
+Post-Training Quantization (PTQ)
 --------------------------------
 
-PTQ enables deploying a model in a low-precision format -- FP8, INT4 or INT8 -- for efficient serving. Different quantization methods are available including FP8 quantization, INT8 SmoothQuant and INT4 AWQ.
+PTQ enables deploying a model in a low-precision format -- FP8, INT4, or INT8 -- for efficient serving. Different quantization methods are available including FP8 quantization, INT8 SmoothQuant, and INT4 AWQ.
 
 Model quantization has two primary benefits: reduced model memory requirements and increased inference throughput.
 
 In NeMo, quantization is enabled by the Nvidia AMMO library -- a unified algorithmic model optimization & deployment toolkit.
 
 The quantization process consists of the following steps:
 
-1. Loading a model checkpoint using appropriate parallelism strategy for evaluation
+1. Loading a model checkpoint using an appropriate parallelism strategy
 2. Calibrating the model to obtain appropriate algorithm-specific scaling factors
-3. Producing output directory or .qnemo tarball with model config (json), quantized weights (safetensors) and tokenizer config (yaml).
+3. Producing an output directory or .qnemo tarball with model config (json), quantized weights (safetensors) and tokenizer config (yaml).
 
-Loading models requires using AMMO spec defined in `megatron.core.deploy.gpt.model_specs module <https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/deploy/gpt/model_specs.py>`_. Typically the calibration step is lightweight and uses a small dataset to obtain appropriate statistics for scaling tensors. The output directory produced (or a .qnemo tarball) is ready to be used to build a serving engine with the Nvidia TensorRT-LLM library. The engine build step is also soon to be the part of NeMo project and ``nemo.deploy`` and ``nemo.export`` modules, see https://github.com/NVIDIA/NeMo/pull/8690.
+Loading models requires using an AMMO spec defined in `megatron.core.inference.gpt.model_specs.py <https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/inference/gpt/model_specs.py>`_ module. Typically the calibration step is lightweight and uses a small dataset to obtain appropriate statistics for scaling tensors. The output directory produced (or a .qnemo tarball) is ready to be used to build a serving engine with the Nvidia TensorRT-LLM library. The engine build step is also available in NeMo project in ``nemo.deploy`` and ``nemo.export`` modules.
 
 Quantization algorithm can also be conveniently set to ``"null"`` to perform only the weights export step using default precision for TensorRT-LLM deployment. This is useful to obtain baseline performance and accuracy results for comparison.
 
 
 Example
 ^^^^^^^
-The example below shows how to quantize the Llama2 70b model into FP8 precision, using tensor parallelism of 8 on a single DGX H100 node. The quantized model is intended for serving using 2 GPUs specified with ``export.inference_tensor_parallel`` parameter.
+The example below shows how to quantize the Llama2 70b model into FP8 precision, using tensor parallelism of 8 on a single DGX H100 node. The quantized model is designed for serving using 2 GPUs specified with the ``export.inference_tensor_parallel`` parameter.
 
-The script should be launched correctly with the number of processes equal to tensor parallelism. This is achieved with the ``mpirun`` command below.
+The script must be launched correctly with the number of processes equal to tensor parallelism. This is achieved with the ``torchrun`` command below:
 
 .. code-block:: bash
 
-    mpirun -n 8 python examples/nlp/language_modeling/megatron_llama_quantization.py \
+    torchrun --nproc-per-node 8 examples/nlp/language_modeling/megatron_llama_quantization.py \
         model_file=llama2-70b-base-bf16.nemo \
         tensor_model_parallel_size=8 \
         pipeline_model_parallel_size=1 \
@@ -57,31 +57,47 @@ The output directory stores the following files:
     └── tokenizer_config.yaml
 
 
-The TensorRT-LLM engine can be build with ``trtllm-build`` command, see `TensorRT-LLM documentation <https://github.com/NVIDIA/TensorRT-LLM/tree/main/examples/llama#fp8-post-training-quantization>`_.
+The TensorRT-LLM engine can be conveniently built and run using ``TensorRTLLM`` class available in ``nemo.export`` submodule:
+
+.. code-block:: python
+
+    from nemo.export import TensorRTLLM
+
+
+    trt_llm_exporter = TensorRTLLM(model_dir="/path/to/trt_llm_engine_folder")
+    trt_llm_exporter.export(
+        nemo_checkpoint_path="llama2-70b-base-fp8-qnemo",
+        model_type="llama",
+    )
+    trt_llm_exporter.forward(["Hi, how are you?", "I am good, thanks, how about you?"])
+
+
+Alternatively, it can also be built directly using ``trtllm-build`` command, see `TensorRT-LLM documentation <https://github.com/NVIDIA/TensorRT-LLM/tree/main/examples/llama#fp8-post-training-quantization>`_:
 
 .. code-block:: bash
 
     trtllm-build \
         --checkpoint_dir llama2-70b-base-fp8-qnemo \
-        --output_dir engine_dir \
+        --output_dir /path/to/trt_llm_engine_folder \
         --max_batch_size 8 \
         --max_input_len 2048 \
-        --max_output_len 512
-
+        --max_output_len 512 \
+        --strongly_typed
 
 
 Known issues
 ^^^^^^^^^^^^
-* Currently in NeMo quantizing and building TensorRT-LLM engines is limited to single-node use cases.
-* Supported and tested model family is Llama2. Quantizing other model types is experimental and may not be fully supported.
-* For INT8 SmoothQuant ``quantization.algorithm=int8_sq``, the TensorRT-LLM engine cannot be build with CLI ``trtllm-build`` command -- Python API and ``tensorrt_llm.builder`` should be used instead.
+* Currently in NeMo, quantizing and building TensorRT-LLM engines is limited to single-node use cases.
+* The supported and tested model family is Llama2. Quantizing other model types is experimental and may not be fully supported.
 
 
 Please refer to the following papers for more details on quantization techniques.
 
 References
 ----------
 
+`Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation, 2020 <https://arxiv.org/abs/2004.09602>`_
+
 `FP8 Formats for Deep Learning, 2022 <https://arxiv.org/abs/2209.05433>`_
 
 `SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models, 2022 <https://arxiv.org/abs/2211.10438>`_

examples/nlp/language_modeling/conf/megatron_llama_quantization.yaml

@@ -25,12 +25,13 @@ quantization:
   algorithm: fp8 # int8_sq, fp8, int8, int4_awq, null
   calib_dataset: cnn_dailymail # wikitext, cnn_dailymail, or a local dataset
   num_calib_size: 512 # number of samples used for calibration
+  awq_block_size: 128 # block size for scaling factors in AWQ algorithm
 
 export:
   decoder_type: llama # gptnext, gpt2, llama
   inference_tensor_parallel: 1 # Default using 1 TP for inference
+  inference_pipeline_parallel: 1 # Default using 1 PP for inference
   dtype: 16 # Default precision data type
-  export_tensorrt_llm_config: true # export config to build TRT-LLM engine directly
 
 model_file: llama2-7b-fp16.nemo # Nemo file path
 model_save: llama2-7b-fp8.qnemo # Path where the quantized model will be saved

examples/nlp/language_modeling/conf/megatron_retro_inference.yaml

@@ -3,42 +3,40 @@ inference:
   top_k: 0  # The number of highest probability vocabulary tokens to keep for top-k-filtering.
   top_p: 0.9 # If set to float < 1, only the most probable tokens with probabilities that add up to top_p or higher are kept for generation.
   temperature: 1.0 # sampling temperature
-  add_BOS: True # add the bos token at the begining of the prompt
+  add_BOS: False # add the bos token at the begining of the prompt
   tokens_to_generate: 30 # The minimum length of the sequence to be generated.
   all_probs: False  # whether return the log prob for all the tokens in vocab
   repetition_penalty: 1.2  # The parameter for repetition penalty. 1.0 means no penalty.
   min_tokens_to_generate: 0  # The minimum length of the sequence to be generated.
   compute_logprob: False  # a flag used to compute logprob of all the input text, a very special case of running inference, default False
-
+  end_strings: ["<|endoftext|>"]  # generation will stop when one of these tokens is generated
+  # RETRO-specific arguments
+  retro_inference:
+    retro_gpt_retrieved_length: 128
+    retro_num_neighbors: 2
+    ft_neighbours: 0
+    reuse_top: False
 
 trainer:
   devices: 1
   num_nodes: 1
   accelerator: gpu
   logger: False # logger provided by exp_manager
-  precision: 16 # 16, 32, or bf16
-
-inference_batch_size: 2
+  precision: 32 # 16, 32, or bf16
+  use_distributed_sampler: False
+  
 tensor_model_parallel_size: -1
 pipeline_model_parallel_size: -1
 pipeline_model_parallel_split_rank: -1 # used for encoder and decoder model (0 for others)
-retro_model_file: null  # RETRO nemo file path
+megatron_amp_O2: False  # Enable O2-level automatic mixed precision to save memory
 
-use_predict_method: False  # whether to use the predict method
+retro_model_file: null  # Retro nemo file path
+checkpoint_dir: null # checkpoint file dir. This is used to load the PTL checkpoint generated during the Retro training
+checkpoint_name: null # PTL checkpoint file name, only used for PTL checkpoint loading
+hparams_file: null # model configuration file, only used for PTL checkpoint loading
 
-prompts: # prompts for RETRO model inference
-  - "hello,"
-  - "good morning,"
-  - "good afternoon,"
-  - "good evening,"
-
-########### Faiss service parameters ########
-retrieval_service:
-  strategy: RetroModelTextGenerationStrategy  # choose customized inference strategy 
-  neighbors: 4
-  frequent_query: False  # for the current token generation, frequently update the retrieval context. If false, update it every 64 tokens 
-  pad_tokens: True # pad the tokens at the beginning to make it minimum of 64 tokens for retrieving at least once
-  store_retrieved: False # whether store the retrieved documents, so it can be checked
-  combo_service:
-    service_ip: '0.0.0.0'
-    service_port: 17181 
+# RETRO inference
+prompt: "sample prompt"
+neighbors:
+  - "neighbor text 1"
+  - "neighbor text 2"

examples/nlp/language_modeling/conf/megatron_retro_inference_legacy.yaml

@@ -0,0 +1,46 @@
+# (This inferencing script for native NeMo RETRO will be soon deprecated. For new inferencing script for mcore RETRO, see ./megatron_retro_inference.yaml)
+
+inference:
+  greedy: False # Whether or not to use sampling ; use greedy decoding otherwise
+  top_k: 0  # The number of highest probability vocabulary tokens to keep for top-k-filtering.
+  top_p: 0.9 # If set to float < 1, only the most probable tokens with probabilities that add up to top_p or higher are kept for generation.
+  temperature: 1.0 # sampling temperature
+  add_BOS: True # add the bos token at the begining of the prompt
+  tokens_to_generate: 30 # The minimum length of the sequence to be generated.
+  all_probs: False  # whether return the log prob for all the tokens in vocab
+  repetition_penalty: 1.2  # The parameter for repetition penalty. 1.0 means no penalty.
+  min_tokens_to_generate: 0  # The minimum length of the sequence to be generated.
+  compute_logprob: False  # a flag used to compute logprob of all the input text, a very special case of running inference, default False
+
+
+trainer:
+  devices: 1
+  num_nodes: 1
+  accelerator: gpu
+  logger: False # logger provided by exp_manager
+  precision: 16 # 16, 32, or bf16
+
+inference_batch_size: 2
+tensor_model_parallel_size: -1
+pipeline_model_parallel_size: -1
+pipeline_model_parallel_split_rank: -1 # used for encoder and decoder model (0 for others)
+retro_model_file: null  # RETRO nemo file path
+
+use_predict_method: False  # whether to use the predict method
+
+prompts: # prompts for RETRO model inference
+  - "hello,"
+  - "good morning,"
+  - "good afternoon,"
+  - "good evening,"
+
+########### Faiss service parameters ########
+retrieval_service:
+  strategy: RetroModelTextGenerationStrategy  # choose customized inference strategy 
+  neighbors: 4
+  frequent_query: False  # for the current token generation, frequently update the retrieval context. If false, update it every 64 tokens 
+  pad_tokens: True # pad the tokens at the beginning to make it minimum of 64 tokens for retrieving at least once
+  store_retrieved: False # whether store the retrieved documents, so it can be checked
+  combo_service:
+    service_ip: '0.0.0.0'
+    service_port: 17181 

examples/nlp/language_modeling/conf/megatron_retro_qatask.yaml

@@ -0,0 +1,40 @@
+inference:
+  greedy: False # Whether or not to use sampling ; use greedy decoding otherwise
+  top_k: 0  # The number of highest probability vocabulary tokens to keep for top-k-filtering.
+  top_p: 0.9 # If set to float < 1, only the most probable tokens with probabilities that add up to top_p or higher are kept for generation.
+  temperature: 1.0 # sampling temperature
+  add_BOS: False # add the bos token at the begining of the prompt
+  tokens_to_generate: 30 # The minimum length of the sequence to be generated.
+  all_probs: False  # whether return the log prob for all the tokens in vocab
+  repetition_penalty: 1.2  # The parameter for repetition penalty. 1.0 means no penalty.
+  min_tokens_to_generate: 0  # The minimum length of the sequence to be generated.
+  compute_logprob: False  # a flag used to compute logprob of all the input text, a very special case of running inference, default False
+  end_strings: ["<|endoftext|>"]  # generation will stop when one of these tokens is generated
+  # RETRO-specific arguments
+  retro_inference:
+    retro_gpt_retrieved_length: 128
+    retro_num_neighbors: 2
+    ft_neighbours: 0
+    reuse_top: False
+
+trainer:
+  devices: 1
+  num_nodes: 1
+  accelerator: gpu
+  logger: False # logger provided by exp_manager
+  precision: 32 # 16, 32, or bf16
+  use_distributed_sampler: False
+
+tensor_model_parallel_size: -1
+pipeline_model_parallel_size: -1
+pipeline_model_parallel_split_rank: -1 # used for encoder and decoder model (0 for others)
+megatron_amp_O2: False  # Enable O2-level automatic mixed precision to save memory
+
+retro_model_file: null  # Retro nemo file path
+checkpoint_dir: null # checkpoint file dir. This is used to load the PTL checkpoint generated during the Retro training
+checkpoint_name: null # PTL checkpoint file name, only used for PTL checkpoint loading
+hparams_file: null # model configuration file, only used for PTL checkpoint loading
+
+# qa tasks
+qa_file_path: null
+pred_file_path: null