Merge branch 'main' into 'main'

megatron升级v0.10

See merge request !3

examples/export/ptq_and_trtllm_export/ptq_trtllm_llama2_7b.sh

+#!/bin/bash
+set -e
+
+DEFAULT_NAME="/checkpoints/llama2-text-7b_v0.2.0"
+NAME="${1:-$DEFAULT_NAME}"
+
+DEFAULT_QUANT_CFG="int8_sq"
+QUANT_CFG="${2:-$DEFAULT_QUANT_CFG}"
+
+# CHANGE THE FOLLOWING IF YOU MOUNT YOUR DATA AND CHECKPOINTS DIFFERENTLY IN THE CONTAINER.
+TP="8"
+INFERENCE_TP=${TP}
+DECODER_TYPE="llama"
+CHECKPOINT_LOAD_DIR="${NAME}"
+TOKENIZER_MODEL="${CHECKPOINT_LOAD_DIR}/hf/tokenizer.model"
+
+# LLaMA2 text 7b has ffn_hidden_size 11008. int4_awq requires a block_size of 128 as a result the TP can at most be 2
+if [ "$QUANT_CFG" = "int4_awq" ]; then
+    INFERENCE_TP="2"
+fi
+
+additional_options=" \
+    --export-quant-cfg ${QUANT_CFG} \
+    --export-legacy-megatron \
+    --export-te-mcore-model \
+    --calib-batch-size 8 \
+    --decoder ${DECODER_TYPE} \
+    --export-dir /tmp/trtllm_ckpt \
+    --inference-tensor-parallel ${INFERENCE_TP} "
+
+trtllm_options=" \
+    --tensorrt-llm-checkpoint-dir /tmp/trtllm_ckpt \
+    --engine-dir /tmp/trtllm_engine \
+    --tokenizer ${CHECKPOINT_LOAD_DIR}/hf \
+    --max-input-len 2048 \
+    --max-output-len 512 \
+    --max-batch-size 8 "
+
+# DO NOT CHANGE THE SETTING BELOW UNLESS YOU KNOW WHAT YOU ARE DOING!!!
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+options=" \
+    --disable-bias-linear \
+    --swiglu \
+    --no-rope-fusion \
+    --untie-embeddings-and-output-weights \
+    --use-rotary-position-embeddings \
+    --normalization RMSNorm \
+    --rotary-percent 1.0 \
+    --no-position-embedding \
+    --no-masked-softmax-fusion \
+    --no-bias-gelu-fusion \
+    --no-bias-dropout-fusion \
+    --no-async-tensor-model-parallel-allreduce \
+    --tensor-model-parallel-size ${TP} \
+    --pipeline-model-parallel-size 1 \
+    --num-layers 32 \
+    --hidden-size 4096 \
+    --ffn-hidden-size 11008 \
+    --num-attention-heads 32 \
+    --seq-length 4096 \
+    --max-position-embeddings 4096 \
+    --micro-batch-size 1 \
+    --make-vocab-size-divisible-by 1 \
+    --tokenizer-type Llama2Tokenizer \
+    --tokenizer-model ${TOKENIZER_MODEL} \
+    --save-interval 1000000 \
+    --use-dist-ckpt \
+    --load ${CHECKPOINT_LOAD_DIR} \
+    --fp16"
+
+# Precompile CUDA extentions
+python -c "import modelopt.torch.quantization.extensions as ext; print(ext.cuda_ext); print(ext.cuda_ext_fp8)"
+
+# Acquire launch configuration where variable launch_config will be set
+launch_config="--nproc_per_node=${TP}"
+
+# Launch multi-process with torchrun
+torchrun ${launch_config} examples/export/ptq_and_trtllm_export/text_generation_ptq.py ${options} ${additional_options}
+

examples/export/ptq_and_trtllm_export/ptq_trtllm_llama3_1_8b.sh

+#!/bin/bash
+set -e
+
+DEFAULT_NAME="/checkpoints/llama-3_1-8b-nemo_v1.0"
+NAME="${1:-$DEFAULT_NAME}"
+
+DEFAULT_QUANT_CFG="int8_sq"
+QUANT_CFG="${2:-$DEFAULT_QUANT_CFG}"
+
+# CHANGE THE FOLLOWING IF YOU MOUNT YOUR DATA AND CHECKPOINTS DIFFERENTLY IN THE CONTAINER.
+TP="1"
+INFERENCE_TP=${TP}
+DECODER_TYPE="llama"
+CHECKPOINT_LOAD_DIR="${NAME}"
+
+# LLaMA2 text 7b has ffn_hidden_size 11008. int4_awq requires a block_size of 128 as a result the TP can at most be 2
+if [ "$QUANT_CFG" = "int4_awq" ]; then
+    INFERENCE_TP="2"
+fi
+
+additional_options=" \
+    --export-quant-cfg ${QUANT_CFG} \
+    --export-legacy-megatron \
+    --export-te-mcore-model \
+    --calib-batch-size 8 \
+    --decoder ${DECODER_TYPE} \
+    --export-dir /tmp/trtllm_ckpt \
+    --inference-tensor-parallel ${INFERENCE_TP} "
+
+# DO NOT CHANGE THE SETTING BELOW UNLESS YOU KNOW WHAT YOU ARE DOING!!!
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+options=" \
+    --disable-bias-linear \
+    --attention-backend unfused \
+    --swiglu \
+    --no-rope-fusion \
+    --untie-embeddings-and-output-weights \
+    --use-rotary-position-embeddings \
+    --normalization RMSNorm \
+    --rotary-percent 1.0 \
+    --hidden-dropout 0.0 \
+    --attention-dropout 0.0 \
+    --no-bias-gelu-fusion \
+    --no-bias-dropout-fusion \
+    --no-async-tensor-model-parallel-allreduce \
+    --tensor-model-parallel-size ${TP} \
+    --pipeline-model-parallel-size 1 \
+    --num-layers 32 \
+    --hidden-size 4096 \
+    --group-query-attention \
+    --num-query-groups 8 \
+    --ffn-hidden-size 14336 \
+    --num-attention-heads 32 \
+    --seq-length 131072 \
+    --max-position-embeddings 131072 \
+    --micro-batch-size 4 \
+    --make-vocab-size-divisible-by 128 \
+    --tokenizer-type HuggingFaceTokenizer \
+    --tokenizer-model meta-llama/Meta-Llama-3.1-8B \
+    --save-interval 1000000 \
+    --use-rope-scaling \
+    --use-dist-ckpt \
+    --load ${CHECKPOINT_LOAD_DIR} \
+    --rotary-base 500000 \
+    --fp16"
+
+# Precompile CUDA extentions
+python -c "import modelopt.torch.quantization.extensions as ext; print(ext.cuda_ext); print(ext.cuda_ext_fp8)"
+
+# Acquire launch configuration where variable launch_config will be set
+launch_config="--nproc_per_node=${TP}"
+
+# Launch multi-process with torchrun
+torchrun ${launch_config} examples/export/ptq_and_trtllm_export/text_generation_ptq.py ${options} ${additional_options}

examples/export/ptq_and_trtllm_export/ptq_trtllm_llama3_8b.sh

+#!/bin/bash
+set -e
+
+DEFAULT_NAME="/checkpoints/llama-3_1-8b-nemo_v1.0"
+NAME="${1:-$DEFAULT_NAME}"
+
+DEFAULT_QUANT_CFG="int8_sq"
+QUANT_CFG="${2:-$DEFAULT_QUANT_CFG}"
+
+
+# CHANGE THE FOLLOWING IF YOU MOUNT YOUR DATA AND CHECKPOINTS DIFFERENTLY IN THE CONTAINER.
+TP="1"
+INFERENCE_TP=${TP}
+DECODER_TYPE="llama"
+CHECKPOINT_LOAD_DIR="${NAME}"
+
+# LLaMA2 text 7b has ffn_hidden_size 11008. int4_awq requires a block_size of 128 as a result the TP can at most be 2
+if [ "$QUANT_CFG" = "int4_awq" ]; then
+    INFERENCE_TP="2"
+fi
+
+additional_options=" \
+    --export-quant-cfg ${QUANT_CFG} \
+    --export-legacy-megatron \
+    --export-te-mcore-model \
+    --calib-batch-size 8 \
+    --decoder ${DECODER_TYPE} \
+    --export-dir /tmp/trtllm_ckpt \
+    --inference-tensor-parallel ${INFERENCE_TP} "
+
+# DO NOT CHANGE THE SETTING BELOW UNLESS YOU KNOW WHAT YOU ARE DOING!!!
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+options=" \
+    --disable-bias-linear \
+    --attention-backend unfused \
+    --swiglu \
+    --no-rope-fusion \
+    --untie-embeddings-and-output-weights \
+    --use-rotary-position-embeddings \
+    --normalization RMSNorm \
+    --rotary-percent 1.0 \
+    --hidden-dropout 0.0 \
+    --attention-dropout 0.0 \
+    --no-bias-gelu-fusion \
+    --no-bias-dropout-fusion \
+    --no-async-tensor-model-parallel-allreduce \
+    --tensor-model-parallel-size ${TP} \
+    --pipeline-model-parallel-size 1 \
+    --num-layers 32 \
+    --hidden-size 4096 \
+    --group-query-attention \
+    --num-query-groups 8 \
+    --ffn-hidden-size 14336 \
+    --num-attention-heads 32 \
+    --seq-length 8192 \
+    --max-position-embeddings 8192 \
+    --micro-batch-size 4 \
+    --make-vocab-size-divisible-by 128 \
+    --tokenizer-type HuggingFaceTokenizer \
+    --tokenizer-model meta-llama/Meta-Llama-3-8B \
+    --save-interval 1000000 \
+    --use-dist-ckpt \
+    --load ${CHECKPOINT_LOAD_DIR} \
+    --rotary-base 500000 \
+    --fp16"
+
+# Precompile CUDA extentions
+python -c "import modelopt.torch.quantization.extensions as ext; print(ext.cuda_ext); print(ext.cuda_ext_fp8)"
+
+# Acquire launch configuration where variable launch_config will be set
+launch_config="--nproc_per_node=${TP}"
+
+# Launch multi-process with torchrun
+torchrun ${launch_config} examples/export/ptq_and_trtllm_export/text_generation_ptq.py ${options} ${additional_options}

examples/export/ptq_and_trtllm_export/ptq_trtllm_minitron_8b.sh

+#!/bin/bash
+set -e
+
+DEFAULT_NAME="/checkpoints/nemotron3-8b_v0.3.0"
+NAME="${1:-$DEFAULT_NAME}"
+
+DEFAULT_QUANT_CFG="fp8"
+QUANT_CFG="${2:-$DEFAULT_QUANT_CFG}"
+
+# CHANGE THE FOLLOWING IF YOU MOUNT YOUR DATA AND CHECKPOINTS DIFFERENTLY IN THE CONTAINER.
+TP="8"
+INFERENCE_TP=${TP}
+DECODER_TYPE="gptnext"
+CHECKPOINT_LOAD_DIR="${NAME}/nemo"
+
+if [ "$QUANT_CFG" = "int4_awq" ]; then
+    INFERENCE_TP="1"
+fi
+
+additional_options=" \
+    --export-quant-cfg ${QUANT_CFG} \
+    --export-legacy-megatron \
+    --export-te-mcore-model \
+    --calib-batch-size 8 \
+    --decoder ${DECODER_TYPE} \
+    --export-dir /tmp/trtllm_ckpt \
+    --inference-tensor-parallel ${INFERENCE_TP} "
+
+# DO NOT CHANGE THE SETTING BELOW UNLESS YOU KNOW WHAT YOU ARE DOING!!!
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+options=" \
+    --apply-layernorm-1p \
+    --attn-attention unfused \
+    --untie-embeddings-and-output-weights \
+    --disable-bias-linear \
+    --no-rope-fusion \
+    --no-position-embedding \
+    --use-rotary-position-embeddings \
+    --rotary-percent 0.5 \
+    --squared-relu \
+    --attention-dropout 0.0 \
+    --hidden-dropout 0.0 \
+    --tensor-model-parallel-size ${TP} \
+    --pipeline-model-parallel-size 1 \
+    --num-layers 32 \
+    --hidden-size 4096 \
+    --ffn-hidden-size 16384 \
+    --group-query-attention \
+    --num-attention-heads 48 \
+    --kv-channels 128 \
+    --seq-length 4096 \
+    --num-query-groups 8 \
+    --max-position-embeddings 4096 \
+    --micro-batch-size 4 \
+    --tokenizer-type HuggingFaceTokenizer \
+    --tokenizer-model nvidia/Minitron-8B-Base \
+    --save-interval 1000000 \
+    --load ${CHECKPOINT_LOAD_DIR} \
+    --bf16 \
+    --use-dist-ckpt"
+
+# Precompile CUDA extentions
+python -c "import modelopt.torch.quantization.extensions as ext; print(ext.cuda_ext); print(ext.cuda_ext_fp8)"
+
+# Acquire launch configuration where variable launch_config will be set
+launch_config="--nproc_per_node=${TP}"
+
+# Launch multi-process with torchrun
+torchrun ${launch_config} examples/export/ptq_and_trtllm_export/text_generation_ptq.py ${options} ${additional_options}

examples/export/ptq_and_trtllm_export/ptq_trtllm_mistral_12b.sh

+#!/bin/bash
+set -e
+
+DEFAULT_NAME="/checkpoints/Mistral-NeMo-12B-Base"
+NAME="${1:-$DEFAULT_NAME}"
+
+DEFAULT_QUANT_CFG="fp8"
+QUANT_CFG="${2:-$DEFAULT_QUANT_CFG}"
+
+# CHANGE THE FOLLOWING IF YOU MOUNT YOUR DATA AND CHECKPOINTS DIFFERENTLY IN THE CONTAINER.
+TP="8"
+INFERENCE_TP=${TP}
+DECODER_TYPE="llama"
+CHECKPOINT_LOAD_DIR="${NAME}"
+
+if [ "$QUANT_CFG" = "int4_awq" ]; then
+    INFERENCE_TP="1"
+fi
+
+additional_options=" \
+    --export-quant-cfg ${QUANT_CFG} \
+    --export-legacy-megatron \
+    --export-te-mcore-model \
+    --calib-batch-size 8 \
+    --decoder ${DECODER_TYPE} \
+    --export-dir /tmp/trtllm_ckpt \
+    --inference-tensor-parallel ${INFERENCE_TP} "
+
+# DO NOT CHANGE THE SETTING BELOW UNLESS YOU KNOW WHAT YOU ARE DOING!!!
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+options=" \
+    --untie-embeddings-and-output-weights \
+    --attention-backend unfused \
+    --disable-bias-linear \
+    --use-rotary-position-embeddings \
+    --rotary-percent 1.0 \
+    --attention-dropout 0.0 \
+    --hidden-dropout 0.0 \
+    --tensor-model-parallel-size ${TP} \
+    --pipeline-model-parallel-size 1 \
+    --num-layers 40 \
+    --hidden-size 5120 \
+    --ffn-hidden-size 14336 \
+    --num-attention-heads 32 \
+    --seq-length 8192 \
+    --kv-channels 128 \
+    --normalization RMSNorm \
+    --swiglu \
+    --num-query-groups 8 \
+    --group-query-attention \
+    --position-embedding-type rope \
+    --max-position-embeddings 8192 \
+    --micro-batch-size 1 \
+    --tokenizer-type HuggingFaceTokenizer \
+    --tiktoken-pattern v2 \
+    --tokenizer-model mistralai/Mistral-Nemo-Base-2407 \
+    --save-interval 1000000 \
+    --load ${CHECKPOINT_LOAD_DIR} \
+    --fp16 \
+    --rotary-base 1000000 \
+    --use-dist-ckpt"
+
+# Precompile CUDA extentions
+python -c "import modelopt.torch.quantization.extensions as ext; print(ext.cuda_ext); print(ext.cuda_ext_fp8)"
+
+# Acquire launch configuration where variable launch_config will be set
+launch_config="--nproc_per_node=${TP}"
+
+# Launch multi-process with torchrun
+torchrun ${launch_config} examples/export/ptq_and_trtllm_export/text_generation_ptq.py ${options} ${additional_options}

examples/export/ptq_and_trtllm_export/ptq_trtllm_mixtral_8x7b.sh

+#!/bin/bash
+set -e
+
+DEFAULT_NAME="/checkpoints/Mistral-NeMo-12B-Base"
+NAME="${1:-$DEFAULT_NAME}"
+
+DEFAULT_QUANT_CFG="fp8"
+QUANT_CFG="${2:-$DEFAULT_QUANT_CFG}"
+
+# NOTE: UNFUSED ATTENTION MUST BE USED TO AVOID ADDITIONAL STATE_DICT KEY MISMATCH.
+export NVTE_FLASH_ATTN=0
+export NVTE_FUSED_ATTN=0
+export NVTE_UNFUSED_ATTN=1
+
+# CHANGE THE FOLLOWING IF YOU MOUNT YOUR DATA AND CHECKPOINTS DIFFERENTLY IN THE CONTAINER.
+TP="8"
+INFERENCE_TP=${TP}
+DECODER_TYPE="llama"
+CHECKPOINT_LOAD_DIR="${NAME}"
+
+if [ "$QUANT_CFG" = "int4_awq" ]; then
+    INFERENCE_TP="1"
+fi
+
+additional_options=" \
+    --export-quant-cfg ${QUANT_CFG} \
+    --export-legacy-megatron \
+    --export-te-mcore-model \
+    --calib-batch-size 8 \
+    --decoder ${DECODER_TYPE} \
+    --export-dir /tmp/trtllm_ckpt \
+    --inference-tensor-parallel ${INFERENCE_TP} "
+
+# DO NOT CHANGE THE SETTING BELOW UNLESS YOU KNOW WHAT YOU ARE DOING!!!
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+options=" \
+    --untie-embeddings-and-output-weights \
+    --no-masked-softmax-fusion \
+    --no-position-embedding \
+    --use-mcore-models \
+    --disable-bias-linear \
+    --rotary-percent 1.0 \
+    --attention-dropout 0.0 \
+    --hidden-dropout 0.0 \
+    --tensor-model-parallel-size ${TP} \
+    --pipeline-model-parallel-size 1 \
+    --num-layers 32 \
+    --hidden-size 4096 \
+    --ffn-hidden-size 14336 \
+    --num-attention-heads 32 \
+    --seq-length 4096 \
+    --kv-channels 128 \
+    --normalization RMSNorm \
+    --swiglu \
+    --num-query-groups 8 \
+    --num-experts 8 \
+    --moe-router-topk 2 \
+    --moe-aux-loss-coeff 1e-2 \
+    --moe-router-load-balancing-type aux_loss \
+    --group-query-attention \
+    --position-embedding-type rope \
+    --no-rope-fusion \
+    --max-position-embeddings 32768 \
+    --micro-batch-size 1 \
+    --tokenizer-type HuggingFaceTokenizer \
+    --tiktoken-pattern v2 \
+    --tokenizer-model mistralai/Mixtral-8x7B-Instruct-v0.1 \
+    --save-interval 1000000 \
+    --load ${CHECKPOINT_LOAD_DIR} \
+    --bf16 \
+    --rotary-base 1000000 \
+    --use-dist-ckpt"
+
+# Precompile CUDA extentions
+python -c "import modelopt.torch.quantization.extensions as ext; print(ext.cuda_ext); print(ext.cuda_ext_fp8)"
+
+# Acquire launch configuration where variable launch_config will be set
+launch_config="--nproc_per_node=${TP}"
+
+# Launch multi-process with torchrun
+torchrun ${launch_config} examples/export/ptq_and_trtllm_export/text_generation_ptq.py ${options} ${additional_options}
+
+

examples/export/ptq_and_trtllm_export/text_generation_ptq.py

+# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
+
+"""Sample Generate GPT."""
+import functools
+import os
+import sys
+from pathlib import Path
+
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../")))
+
+import modelopt.torch.quantization as mtq
+import torch
+from datasets import load_dataset
+from tqdm import tqdm
+
+# [ModelOpt]: changing the default model provider to the ModelOpt version
+from megatron.core import mpu
+from megatron.inference.arguments import add_modelopt_args
+from megatron.inference.checkpointing import load_modelopt_checkpoint
+from megatron.inference.gpt.model_provider import model_provider
+from megatron.inference.text_generation import generate_and_post_process
+from megatron.training import get_args, get_model, initialize_megatron
+from megatron.training.checkpointing import save_checkpoint
+from megatron.training.utils import print_rank_0, unwrap_model
+
+QUANT_CFG_CHOICES = {
+    "int8": mtq.INT8_DEFAULT_CFG,
+    "int8_sq": mtq.INT8_SMOOTHQUANT_CFG,
+    "fp8": mtq.FP8_DEFAULT_CFG,
+    "int4_awq": mtq.INT4_AWQ_CFG,
+    "w4a8_awq": mtq.W4A8_AWQ_BETA_CFG,
+    "int4": mtq.INT4_BLOCKWISE_WEIGHT_ONLY_CFG,
+}
+
+
+def add_trtllm_ckpt_export_args(parser):
+    """Add additional arguments for TensorRT-LLM."""
+    group = parser.add_argument_group(title="trtllm")
+
+    group.add_argument(
+        "--export-dir", type=str, help="The output TensorRT-LLM checkpoint.",
+    )
+    group.add_argument(
+        "--decoder", type=str, choices=["gptnext", 'llama'], help="The decoder type of the model.",
+    )
+    group.add_argument(
+        "--inference-tensor-parallel",
+        type=int,
+        help="Tensor parallel for the inference time, can be different from the training config.",
+        default=1,
+    )
+
+
+def add_text_generate_ptq_args(parser):
+    """Add additional arguments for ModelOpt text generation PTQ."""
+    group = parser.add_argument_group(title='ModelOpt text generation ptq')
+    group.add_argument(
+        "--calib-dataset",
+        type=str,
+        default="cnn_dailymail",
+        help="Calibration datasets from HuggingFace datasets.",
+    )
+    group.add_argument(
+        "--calib-batch-size", type=int, default=4, help="Batch size to use for ptq calibration."
+    )
+    group.add_argument(
+        "--calib-size", type=int, default=512, help="Samples to use for ptq calibration."
+    )
+    parser.add_argument(
+        "--prompts",
+        type=str,
+        default=(
+            "Born in north-east France, Soyer trained as a|Born in California, Soyer trained as a"
+        ),
+        help="Input texts. Please use | to separate different batches.",
+    )
+    add_modelopt_args(parser)
+    add_trtllm_ckpt_export_args(parser)
+    return parser
+
+
+def get_calib_dataloader(
+    data="cnn_dailymail", batch_size=4, calib_size=512, max_sequence_length=512
+):
+    if data == "pileval":
+        dataset = load_dataset(
+            "json", data_files="https://the-eye.eu/public/AI/pile/val.jsonl.zst", split="train"
+        )
+        text_column = "text"
+    elif data == "wikitext":
+        dataset = load_dataset("wikitext", "wikitext-103-v1", split="train")
+        text_column = "text"
+    elif data == "cnn_dailymail":
+        dataset = load_dataset("cnn_dailymail", name="3.0.0", split="train")
+        text_column = "article"
+
+    calib_size = max(min(len(dataset), calib_size), batch_size)
+    for i in range(calib_size // batch_size):
+        batch = dataset[i * batch_size : (i + 1) * batch_size][text_column]
+        for j in range(len(batch)):
+            batch[j] = batch[j][:max_sequence_length]
+        yield batch
+
+
+
+if __name__ == "__main__":
+    initialize_megatron(
+        extra_args_provider=add_text_generate_ptq_args,
+        args_defaults={
+            'tokenizer_type': 'GPT2BPETokenizer',
+            'no_load_rng': True,
+            'no_load_optim': True,
+        },
+    )
+
+    args = get_args()
+    if args.num_layers_per_virtual_pipeline_stage is not None:
+        print_rank_0("Interleaved pipeline schedule is not yet supported for text generation.")
+        exit()
+
+    print_rank_0("WARNING: Forcing exit_on_missing_checkpoint to True for text generation.")
+    args.exit_on_missing_checkpoint = True
+    if hasattr(args, 'moe_grouped_gemm') and args.moe_grouped_gemm == True:
+        print_rank_0("WARNING: Forcing moe_grouped_gemm to False for PTQ and export.")
+        args.moe_grouped_gemm = False
+
+    # Set up model and load checkpoint
+    # [ModelOpt]: make sure that output logits are allgathered.
+    text_generation_model_provider = functools.partial(model_provider, parallel_output=False)
+    model = get_model(text_generation_model_provider, wrap_with_ddp=False)
+
+    if args.load is not None:
+        load_modelopt_checkpoint(model, strict=not args.untie_embeddings_and_output_weights)
+        print_rank_0("Done loading checkpoint")
+
+    # Removing virtual pipeline parallel and other wrapper
+    assert len(model) == 1, "Above condition should have caught this"
+    unwrapped_model = unwrap_model(model)
+
+    all_prompts = args.prompts.split("|")
+
+    def custom_prompt_forward_loop_func(model):
+        for prompt in tqdm(all_prompts):
+            if mpu.is_pipeline_first_stage() and mpu.get_tensor_model_parallel_rank() == 0:
+                (
+                    prompts_plus_generations,
+                    prompts_plus_generations_segments,
+                    logprobs,
+                    _,
+                ) = generate_and_post_process(
+                    model,
+                    prompts=[prompt],
+                    tokens_to_generate=128,
+                    return_output_log_probs=True,
+                    temperature=1.0,
+                )
+                print_rank_0(prompts_plus_generations)
+            else:
+                generate_and_post_process(model)
+
+    def hf_dataset_forword_loop_func(model):
+        dataloader = get_calib_dataloader(args.calib_dataset, args.calib_batch_size, args.calib_size)
+        for prompts in tqdm(dataloader, total=args.calib_size//args.calib_batch_size):
+            if mpu.is_pipeline_first_stage() and mpu.get_tensor_model_parallel_rank() == 0:
+                (
+                    prompts_plus_generations,
+                    prompts_plus_generations_segments,
+                    logprobs,
+                    _,
+                ) = generate_and_post_process(
+                    model,
+                    prompts=prompts,
+                    tokens_to_generate=0,
+                    return_output_log_probs=False,
+                    temperature=1.0,
+                )
+            else:
+                generate_and_post_process(model)
+
+    ptq_forward_loop_func = custom_prompt_forward_loop_func
+    if args.calib_dataset is not None:
+        ptq_forward_loop_func = hf_dataset_forword_loop_func
+
+    if args.export_quant_cfg in QUANT_CFG_CHOICES:
+        mtq_config = QUANT_CFG_CHOICES[args.export_quant_cfg]
+        if "*output_layer*" not in mtq_config["quant_cfg"]:
+            mtq_config["quant_cfg"]["*output_layer*"] = {"enable": False}
+        if "awq" in args.export_quant_cfg:
+            weight_quantizer = mtq_config["quant_cfg"]["*weight_quantizer"]  # type: ignore
+            if isinstance(weight_quantizer, list):
+                weight_quantizer = weight_quantizer[0]
+            weight_quantizer["block_sizes"][-1] = 128
+        print_rank_0("Quantizing the model...")
+        mtq.quantize(unwrapped_model[0], mtq_config, ptq_forward_loop_func)
+
+    custom_prompt_forward_loop_func(model[0])
+
+    if args.save is not None and args.export_quant_cfg in QUANT_CFG_CHOICES:
+        save_checkpoint(1, unwrapped_model, None, None, 0)
+
+    print_rank_0(f"Fake Quantized Model:\n {unwrapped_model[0]}")
+
+    if args.export_dir:
+        assert args.decoder in ["gptnext", "llama"], f"Decoder type {args.decoder} not supported."
+        Path(args.export_dir).mkdir(parents=True, exist_ok=True)
+        print_rank_0("Exporting TensorRT-LLM checkpoints.")
+
+        from modelopt.torch.export import export_tensorrt_llm_checkpoint
+
+        # In TRT LLM, squared relu activation does not support bf16. So we use fp16 by default.
+        export_tensorrt_llm_checkpoint(
+            unwrapped_model[0],
+            args.decoder,
+            torch.bfloat16 if args.bf16 else torch.float16,
+            export_dir=args.export_dir,
+            inference_tensor_parallel=args.inference_tensor_parallel,
+            inference_pipeline_parallel=1,
+            use_nfs_workspace=True,
+        )
+
+        print_rank_0(f"TensorRT-LLM checkpoints saved to {args.export_dir}")
+        torch.distributed.barrier()

examples/export/ptq_and_trtllm_export/trtllm_text_generation.py

+# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
+
+"""An example script to run the tensorrt_llm engine."""
+
+import argparse
+from pathlib import Path
+import subprocess
+from typing import Optional, Union
+
+import numpy as np
+import torch
+from modelopt.deploy.llm import LLM
+from tensorrt_llm.models import PretrainedConfig
+from transformers import AutoTokenizer, T5Tokenizer
+import tensorrt_llm
+
+
+def parse_arguments():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--tokenizer", type=str, default="")
+    parser.add_argument("--engine-dir", type=str, default="/tmp/trtllm_engine")
+    parser.add_argument(
+        "--input-texts",
+        type=str,
+        default=(
+            "Born in north-east France, Soyer trained as a|Born in California, Soyer trained as a"
+        ),
+        help="Input texts. Please use | to separate different batches.",
+    )
+    return parser.parse_args()
+
+
+def run(args):
+    try:
+        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer, trust_remote_code=True)
+    except Exception as e:
+        raise Exception(f"Failed to load tokenizer: {e}")
+
+    print(tokenizer, tokenizer.vocab_size)
+
+    input_texts = args.input_texts.split("|")
+    assert input_texts, "input_text not specified"
+    print(input_texts)
+
+    free_memory_before = torch.cuda.mem_get_info()
+
+    # This is a ModelOpt wrapper on top of tensorrt_llm.hlapi.llm.LLM
+    llm_engine = LLM(args.engine_dir, tokenizer)
+
+    torch.cuda.cudart().cudaProfilerStart()
+    # outputs = llm_engine.generate_text(input_texts, args.max_output_len, args.max_beam_width)
+    outputs = llm_engine.generate(input_texts)
+    torch.cuda.cudart().cudaProfilerStop()
+
+    free_memory_after = torch.cuda.mem_get_info()
+    print(
+        f"Used GPU memory: {(free_memory_before[0] - free_memory_after[0]) / 1024 / 1024 / 1024} GB"
+    )
+    print(outputs)
+
+
+if __name__ == "__main__":
+    args = parse_arguments()
+    run(args)

examples/export/trtllm_export/README.md

+# Megatron Core To TRTLLM Export Documentation
+This guide will walk you through how you can use the megatron core export for exporting models to trtllm format
+
+### Contents
+- [Megatron Core To TRTLLM Export Documentation](#megatron-core-to-trtllm-export-documentation)
+- [Contents](#contents)
+  - [1. Quick Start](#1-quick-start)
+    - [1.1 Understanding The Code](#11-understanding-the-code)
+    - [1.2 Running The Code](#12-running-the-code)
+  - [2. GPU Export](#2-gpu-export)
+  - [3. Future work](#4-future-work)
+
+#### 1. Quick Start
+This will walk you through the flow of converting an mcore gpt model to trtllm format using single device mode. The file can be found at [gpt_single_device_cpu_export.py](./single_device_export/gpt_single_device_cpu_export.py)
+
+NOTE: For faster performance, if your entire model will fit into gpu memory, pre transfer the model state dict to gpu and then call the get_trtllm_pretrained_config_and_model_weights function.
+
+<br>
+
+##### 1.1 Understanding The Code
+***STEP 1 - We initialize model parallel and other default arguments***
+We initalize tp and pp to 1 so that we can get the full model state dict on cpu
+```python
+    initialize_distributed(tensor_model_parallel_size=1, pipeline_model_parallel_size=1)
+```
+
+***STEP 2 - We load the model using the model_provider_function***
+NOTE: We create a simple gpt model
+
+```python
+    transformer_config = TransformerConfig(
+        num_layers=2, 
+        hidden_size=64, # Needs to be atleast 32 times num_attn_heads
+        num_attention_heads=2, 
+        use_cpu_initialization=True, 
+        pipeline_dtype=torch.float32,
+    )
+
+    gpt_model = GPTModel(
+        config=transformer_config, 
+        transformer_layer_spec=get_gpt_layer_local_spec(), 
+        vocab_size=100, 
+        max_sequence_length=_SEQUENCE_LENGTH,
+    )
+
+    # Optionally you can also load a model using this code 
+    # sharded_state_dict=gpt_model.sharded_state_dict(prefix='')
+    # checkpoint = dist_checkpointing.load(sharded_state_dict=sharded_state_dict, checkpoint_dir=checkpoint_path)
+    # gpt_model.load_state_dict(checkpoint)
+
+```
+
+***STEP 3 - Instantiate the TRTLLM Helper***
+We instantiate the [TRTLLM Helper](../../../megatron/core/export/trtllm/trtllm_helper.py)  For the GPT model we instantiate trtllm_helper as shown below.
+```python
+    if hasattr(gpt_model, "rotary_pos_emb"):
+        seq_len_interpolation_factor =  gpt_model.rotary_pos_emb.seq_len_interpolation_factor
+
+    trtllm_helper = TRTLLMHelper(
+                        transformer_config=gpt_model.config, 
+                        model_type=ModelType.gpt,
+                        position_embedding_type = gpt_model.position_embedding_type, 
+                        max_position_embeddings = gpt_model.max_position_embeddings, 
+                        rotary_percentage = gpt_model.rotary_percent,
+                        rotary_base = gpt_model.rotary_base,
+                        moe_tp_mode = 2,
+                        multi_query_mode = False,
+                        activation = "gelu", 
+                        seq_len_interpolation_factor = seq_len_interpolation_factor,
+                        share_embeddings_and_output_weights=gpt_model.share_embeddings_and_output_weights
+                    )   
+```
+
+***STEP 4 - Get the TRTLLM Weights and configs***
+To convert model weights to trtllm weights and configs, we use the [single_device_converter](../../../megatron/core/export/trtllm/trtllm_weights_converter/single_device_trtllm_model_weights_converter.py). We pass as inputs the model state dict, and export config. In this example we use inference tp size as 2 for the export. 
+
+```python
+    model_state_dict={}
+    for key , val in gpt_model.state_dict().items():
+        # val is non for _extra_state layers . We filter it out
+        if val is not None:
+            model_state_dict[key] = val
+
+    export_config = ExportConfig(inference_tp_size = 2)
+    weight_list, config_list = trtllm_helper.get_trtllm_pretrained_config_and_model_weights(
+        model_state_dict= model_state_dict,
+        dtype = DataType.bfloat16,
+        export_config=export_config
+    )
+```
+
+***STEP 5 - Build the TRTLLM Engine***
+Following code is used to build the TRTLLM Engine. 
+
+```python
+    for trtllm_model_weights, trtllm_model_config in zip(weight_list, config_list):
+        trtllm_helper.build_and_save_engine(
+            max_input_len=256,
+            max_output_len=256,
+            max_batch_size=8,
+            engine_dir='/opt/megatron-lm/engine',
+            trtllm_model_weights=trtllm_model_weights,
+            trtllm_model_config=trtllm_model_config,
+            lora_ckpt_list=None,
+            use_lora_plugin=None,
+            max_lora_rank=64,
+            lora_target_modules=None,
+            max_prompt_embedding_table_size=0,
+            paged_kv_cache=True,
+            remove_input_padding=True,
+            paged_context_fmha=False,
+            use_refit=False,
+            max_num_tokens=None,
+            max_seq_len=512,
+            opt_num_tokens=None,
+            max_beam_width=1,
+            tokens_per_block=128,
+            multiple_profiles=False,
+            gpt_attention_plugin="auto",
+            gemm_plugin="auto",
+        )
+```
+<br>
+
+##### 1.2 Running The Code
+An example run script is shown below. 
+
+```
+# In a workstation 
+MLM_PATH=/path/to/megatron-lm
+CONTAINER_IMAGE=gitlab-master.nvidia.com:5005/dl/joc/nemo-ci/trtllm_0.12/train:pipe.17669124-x86
+
+docker run -it --gpus=all --ipc=host -v $MLM_PATH/:/opt/megatron-lm $CONTAINER_IMAGE bash
+
+# Inside the container run the following. 
+
+cd /opt/megatron-lm/
+
+CUDA_VISIBLE_DEVICES=0 torchrun --nproc-per-node 1  examples/export/trtllm_export/single_device_export/gpt_single_device_cpu_export.py
+```
+
+<br>
+
+#### 2. GPU Export
+You can use the [gpt_distributed_gpu_export.py](./distributed_export/gpt_distributed_gpu_export.py) to run a more optimized on device distributed. version of trtllm export. Internally this uses the [distributed_converter](../../../megatron/core/export/trtllm/trtllm_weights_converter/distributed_trtllm_model_weights_converter.py) to convert model weights on device. 
+In the single device version you collect all the model weights on CPU/GPU, convert it to trtllm format, and then store the engine back on disk. In the GPU version you load each individual state dict on the gpus, convert it on the device itself and store the engine on disk. 
+
+To run the gpu version 
+
+```
+CUDA_VISIBLE_DEVICES=0,1 torchrun --nproc-per-node 2  examples/export/trtllm_export/distributed_export/gpt_distributed_gpu_export.py
+```
+
+<br>
+
+#### 3. Future work
+The following are planned for the future releases . 
+* Pipeline parallellism for export (Work in progress) 
+* GPU Export for more models (Work in progress for some models)
+* Refit functionality
+* VLLM Support
\ No newline at end of file

examples/export/trtllm_export/distributed_export/gpt_distributed_gpu_export.py

+import os
+import torch
+from megatron.core import parallel_state
+from megatron.core import dist_checkpointing
+from megatron.core.export.model_type import ModelType
+from megatron.core.export.data_type import DataType
+from megatron.core.export.trtllm.trtllm_helper import TRTLLMHelper
+from megatron.core.tensor_parallel.random import model_parallel_cuda_manual_seed
+from megatron.core.transformer.transformer_config import TransformerConfig
+from megatron.core.models.gpt.gpt_model import GPTModel
+from megatron.core.models.gpt.gpt_layer_specs import get_gpt_layer_local_spec
+
+
+_SEQUENCE_LENGTH = 64
+_VOCAB_SIZE = 256
+
+def initialize_distributed(tensor_model_parallel_size=1, pipeline_model_parallel_size=1):
+    parallel_state.destroy_model_parallel()
+
+    # Torch setup for distributed training
+    rank = int(os.environ['LOCAL_RANK'])
+    world_size = torch.cuda.device_count()
+    torch.cuda.set_device(rank)
+    torch.distributed.init_process_group(world_size=world_size, rank=rank)
+
+    # Megatron core distributed training initialization
+    parallel_state.initialize_model_parallel(tensor_model_parallel_size = tensor_model_parallel_size, pipeline_model_parallel_size=pipeline_model_parallel_size)
+
+def model_provider():
+    """Build the model."""
+
+    transformer_config = TransformerConfig(
+        num_layers=2, 
+        hidden_size=64, 
+        num_attention_heads=2, 
+        use_cpu_initialization=True, 
+        pipeline_dtype=torch.float32
+    )
+
+    gpt_model = GPTModel(
+        config=transformer_config, 
+        transformer_layer_spec=get_gpt_layer_local_spec(), 
+        vocab_size=_VOCAB_SIZE, 
+        max_sequence_length=_SEQUENCE_LENGTH,
+    )
+
+    return gpt_model
+
+def load_distributed_checkpoint(checkpoint_path, gpt_model):
+    sharded_state_dict=gpt_model.sharded_state_dict(prefix='')
+    checkpoint = dist_checkpointing.load(sharded_state_dict=sharded_state_dict, checkpoint_dir=checkpoint_path)
+    gpt_model.load_state_dict(checkpoint)
+    return gpt_model
+
+if __name__ == "__main__":
+    initialize_distributed(tensor_model_parallel_size=2, pipeline_model_parallel_size=1)
+    model_parallel_cuda_manual_seed(123)
+
+    gpt_model = model_provider()
+    device = torch.device("cuda")
+    gpt_model.to(device) 
+    
+    # Optionally you can also load a gpt model from ckpt_path using this code below
+    # gpt_model = load_distributed_checkpoint(gpt_model=gpt_model, checkpoint_path=ckpt_path)
+
+    seq_len_interpolation_factor = None
+    if hasattr(gpt_model, "rotary_pos_emb"):
+        seq_len_interpolation_factor =  gpt_model.rotary_pos_emb.seq_len_interpolation_factor
+
+    trtllm_helper = TRTLLMHelper(
+                        transformer_config=gpt_model.config, 
+                        model_type=ModelType.gpt,
+                        position_embedding_type = gpt_model.position_embedding_type, 
+                        max_position_embeddings = gpt_model.max_position_embeddings, 
+                        rotary_percentage = gpt_model.rotary_percent,
+                        rotary_base = gpt_model.rotary_base,
+                        moe_tp_mode = 2,
+                        multi_query_mode = False,
+                        activation = "gelu", 
+                        seq_len_interpolation_factor = seq_len_interpolation_factor,
+                        share_embeddings_and_output_weights=gpt_model.share_embeddings_and_output_weights
+                    )
+    
+
+    trtllm_model_weights, trtllm_model_config = trtllm_helper.get_trtllm_pretrained_config_and_model_weights(
+        model_state_dict= gpt_model.state_dict(),
+        dtype = DataType.bfloat16,
+        on_device_distributed_conversion=True, 
+        vocab_size=_VOCAB_SIZE, 
+        gpus_per_node=2,
+    )
+
+    trtllm_helper.build_and_save_engine(
+        max_input_len=256,
+        max_output_len=256,
+        max_batch_size=8,
+        engine_dir='/opt/megatron-lm/engine',
+        trtllm_model_weights=trtllm_model_weights[0],
+        trtllm_model_config=trtllm_model_config[0],
+        lora_ckpt_list=None,
+        use_lora_plugin=None,
+        max_lora_rank=64,
+        lora_target_modules=None,
+        max_prompt_embedding_table_size=0,
+        paged_kv_cache=True,
+        remove_input_padding=True,
+        paged_context_fmha=False,
+        use_refit=False,
+        max_num_tokens=None,
+        max_seq_len=512,
+        opt_num_tokens=None,
+        max_beam_width=1,
+        tokens_per_block=128,
+        multiple_profiles=False,
+        gpt_attention_plugin="auto",
+        gemm_plugin="auto",
+    )

examples/export/trtllm_export/single_device_export/gpt_single_device_cpu_export.py

+import os
+import torch
+from megatron.core import parallel_state
+from megatron.core import dist_checkpointing
+from megatron.core.export.model_type import ModelType
+from megatron.core.export.data_type import DataType
+from megatron.core.export.export_config import ExportConfig
+from megatron.core.export.trtllm.trtllm_helper import TRTLLMHelper
+from megatron.core.tensor_parallel.random import model_parallel_cuda_manual_seed
+from megatron.core.transformer.transformer_config import TransformerConfig
+from megatron.core.models.gpt.gpt_model import GPTModel
+from megatron.core.models.gpt.gpt_layer_specs import get_gpt_layer_local_spec
+
+
+_SEQUENCE_LENGTH = 64
+
+
+def initialize_distributed(tensor_model_parallel_size=1, pipeline_model_parallel_size=1):
+    parallel_state.destroy_model_parallel()
+
+    # Torch setup for distributed training
+    rank = int(os.environ['LOCAL_RANK'])
+    world_size = torch.cuda.device_count()
+    torch.cuda.set_device(rank)
+    torch.distributed.init_process_group(world_size=world_size, rank=rank)
+
+    # Megatron core distributed training initialization
+    parallel_state.initialize_model_parallel(tensor_model_parallel_size, pipeline_model_parallel_size)
+
+def model_provider():
+    """Build the model."""
+
+    transformer_config = TransformerConfig(
+        num_layers=2, 
+        hidden_size=64, # Needs to be atleast 32 times num_attn_heads
+        num_attention_heads=2, 
+        use_cpu_initialization=True, 
+        pipeline_dtype=torch.float32,
+    )
+
+    gpt_model = GPTModel(
+        config=transformer_config, 
+        transformer_layer_spec=get_gpt_layer_local_spec(), 
+        vocab_size=100, 
+        max_sequence_length=_SEQUENCE_LENGTH,
+    )
+
+    return gpt_model
+
+def load_distributed_checkpoint(checkpoint_path, gpt_model):
+    sharded_state_dict=gpt_model.sharded_state_dict(prefix='')
+    checkpoint = dist_checkpointing.load(sharded_state_dict=sharded_state_dict, checkpoint_dir=checkpoint_path)
+    gpt_model.load_state_dict(checkpoint)
+    return gpt_model
+
+if __name__ == "__main__":
+    # Need to use TP1 PP1 for export on single device
+    initialize_distributed(tensor_model_parallel_size=1, pipeline_model_parallel_size=1)
+    model_parallel_cuda_manual_seed(123)
+
+    gpt_model = model_provider()
+
+    # Optionally you can also load a gpt model from ckpt_path using this code below
+    # gpt_model = load_distributed_checkpoint(gpt_model=gpt_model, checkpoint_path=ckpt_path)
+
+    seq_len_interpolation_factor = None
+    if hasattr(gpt_model, "rotary_pos_emb"):
+        seq_len_interpolation_factor =  gpt_model.rotary_pos_emb.seq_len_interpolation_factor
+
+    trtllm_helper = TRTLLMHelper(
+                        transformer_config=gpt_model.config, 
+                        model_type=ModelType.gpt,
+                        position_embedding_type = gpt_model.position_embedding_type, 
+                        max_position_embeddings = gpt_model.max_position_embeddings, 
+                        rotary_percentage = gpt_model.rotary_percent,
+                        rotary_base = gpt_model.rotary_base,
+                        moe_tp_mode = 2,
+                        multi_query_mode = False,
+                        activation = "gelu", 
+                        seq_len_interpolation_factor = seq_len_interpolation_factor,
+                        share_embeddings_and_output_weights=gpt_model.share_embeddings_and_output_weights
+                    )
+    
+
+    export_config = ExportConfig(inference_tp_size = 2)
+    # NOTE : For faster performance, if your entire model will fit in gpu memory, transfer model state dict to GPU and then call this api
+    weight_list, config_list = trtllm_helper.get_trtllm_pretrained_config_and_model_weights(
+        model_state_dict= gpt_model.state_dict(),
+        dtype = DataType.bfloat16,
+        export_config=export_config
+    )
+
+    for trtllm_model_weights, trtllm_model_config in zip(weight_list, config_list):
+        trtllm_helper.build_and_save_engine(
+            max_input_len=256,
+            max_output_len=256,
+            max_batch_size=8,
+            engine_dir='/opt/megatron-lm/engine',
+            trtllm_model_weights=trtllm_model_weights,
+            trtllm_model_config=trtllm_model_config,
+            lora_ckpt_list=None,
+            use_lora_plugin=None,
+            max_lora_rank=64,
+            lora_target_modules=None,
+            max_prompt_embedding_table_size=0,
+            paged_kv_cache=True,
+            remove_input_padding=True,
+            paged_context_fmha=False,
+            use_refit=False,
+            max_num_tokens=None,
+            max_seq_len=512,
+            opt_num_tokens=None,
+            max_beam_width=1,
+            tokens_per_block=128,
+            multiple_profiles=False,
+            gpt_attention_plugin="auto",
+            gemm_plugin="auto",
+        )
\ No newline at end of file

examples/finetune_mnli_distributed.sh

-#!/bin/bash
-
-WORLD_SIZE=8
-
-DISTRIBUTED_ARGS="--nproc_per_node $WORLD_SIZE \
-                  --nnodes 1 \
-                  --node_rank 0 \
-                  --master_addr localhost \
-                  --master_port 6000"
-
-TRAIN_DATA="data/glue_data/MNLI/train.tsv"
-VALID_DATA="data/glue_data/MNLI/dev_matched.tsv \
-            data/glue_data/MNLI/dev_mismatched.tsv"
-PRETRAINED_CHECKPOINT=checkpoints/bert_345m
-VOCAB_FILE=bert-vocab.txt
-CHECKPOINT_PATH=checkpoints/bert_345m_mnli
-
-python -m torch.distributed.launch $DISTRIBUTED_ARGS ./tasks/main.py \
-               --task MNLI \
-               --seed 1234 \
-               --train-data $TRAIN_DATA \
-               --valid-data $VALID_DATA \
-               --tokenizer-type BertWordPieceLowerCase \
-               --vocab-file $VOCAB_FILE \
-               --epochs 5 \
-               --pretrained-checkpoint $PRETRAINED_CHECKPOINT \
-               --tensor-model-parallel-size 1 \
-               --num-layers 24 \
-               --hidden-size 1024 \
-               --num-attention-heads 16 \
-               --micro-batch-size 8 \
-               --lr 5.0e-5 \
-               --lr-decay-style linear \
-               --lr-warmup-fraction 0.065 \
-               --seq-length 512 \
-               --max-position-embeddings 512 \
-               --save-interval 500000 \
-               --save $CHECKPOINT_PATH \
-               --log-interval 10 \
-               --eval-interval 100 \
-               --eval-iters 50 \
-               --weight-decay 1.0e-1 \
-               --fp16

examples/finetune_race_distributed.sh

-#!/bin/bash
-
-WORLD_SIZE=8
-
-DISTRIBUTED_ARGS="--nproc_per_node $WORLD_SIZE \
-                  --nnodes 1 \
-                  --node_rank 0 \
-                  --master_addr localhost \
-                  --master_port 6000"
-
-TRAIN_DATA="data/RACE/train/middle"
-VALID_DATA="data/RACE/dev/middle \
-            data/RACE/dev/high"
-VOCAB_FILE=bert-vocab.txt
-PRETRAINED_CHECKPOINT=checkpoints/bert_345m
-CHECKPOINT_PATH=checkpoints/bert_345m_race
-
-python -m torch.distributed.launch $DISTRIBUTED_ARGS ./tasks/main.py \
-               --task RACE \
-               --seed 1234 \
-               --train-data $TRAIN_DATA \
-               --valid-data $VALID_DATA \
-               --tokenizer-type BertWordPieceLowerCase \
-               --vocab-file $VOCAB_FILE \
-               --epochs 3 \
-               --pretrained-checkpoint $PRETRAINED_CHECKPOINT \
-               --tensor-model-parallel-size 1 \
-               --num-layers 24 \
-               --hidden-size 1024 \
-               --num-attention-heads 16 \
-               --micro-batch-size 4 \
-               --lr 1.0e-5 \
-               --lr-decay-style linear \
-               --lr-warmup-fraction 0.06 \
-               --seq-length 512 \
-               --max-position-embeddings 512 \
-               --save-interval 100000 \
-               --save $CHECKPOINT_PATH \
-               --log-interval 10 \
-               --eval-interval 100 \
-               --eval-iters 50 \
-               --weight-decay 1.0e-1 \
-               --clip-grad 1.0 \
-               --hidden-dropout 0.1 \
-               --attention-dropout 0.1 \
-               --fp16

examples/finetune_retriever_distributed.sh

-#!/bin/bash
-
-# Finetune a BERT or pretrained ICT model using Google natural question data 
-# Datasets can be downloaded from the following link:
-# https://github.com/facebookresearch/DPR/blob/master/data/download_data.py
-
-WORLD_SIZE=8
-
-DISTRIBUTED_ARGS="--nproc_per_node $WORLD_SIZE \
-                  --nnodes 1 \
-                  --node_rank 0 \
-                  --master_addr localhost \
-                  --master_port 6000"
-
-CHECKPOINT_PATH=<Specify path for the finetuned retriever model>
-
-# Load either of the below
-BERT_LOAD_PATH=<Path of BERT pretrained model>
-PRETRAINED_CHECKPOINT=<Path of Pretrained ICT model>
-
-python -m torch.distributed.launch $DISTRIBUTED_ARGS ./tasks/main.py \
-        --task RET-FINETUNE-NQ \
-        --train-with-neg \
-        --train-hard-neg 1 \
-        --pretrained-checkpoint ${PRETRAINED_CHECKPOINT} \
-        --num-layers 12 \
-        --hidden-size 768 \
-        --num-attention-heads 12 \
-        --tensor-model-parallel-size 1 \
-        --tokenizer-type BertWordPieceLowerCase \
-        --train-data nq-train.json \
-        --valid-data nq-dev.json \
-        --save ${CHECKPOINT_PATH} \
-        --load ${CHECKPOINT_PATH} \
-        --vocab-file bert-vocab.txt \
-        --bert-load ${BERT_LOAD_PATH} \
-        --save-interval 5000 \
-        --log-interval 10 \
-        --eval-interval 20000 \
-        --eval-iters 100 \
-        --indexer-log-interval 1000 \
-        --faiss-use-gpu \
-        --DDP-impl torch \
-        --fp16 \
-        --retriever-report-topk-accuracies 1 5 10 20 100 \
-        --seq-length 512 \
-        --retriever-seq-length 256 \
-        --max-position-embeddings 512 \
-        --retriever-score-scaling \
-        --epochs 80 \
-        --micro-batch-size 8 \
-        --eval-micro-batch-size 16 \
-        --indexer-batch-size 128 \
-        --lr 2e-5 \
-        --lr-warmup-fraction 0.01 \
-        --weight-decay 1e-1

examples/gpt3/README.md

+# GPT3 MODEL
+
+## Table of contents
+- [1. Training Setup](#1-training-setup)
+- [2. Configurations](#2-configurations)
+- [3. Training Results](#3-training-results)
+
+## 1. Training setup
+<a id="markdown-training-setup" name="training-setup"></a>
+
+To run the model using a docker container run it as follows
+```
+PYTORCH_IMAGE=nvcr.io/nvidia/pytorch:24.01-py3
+CHECKPOINT_PATH="" #<Specify path>
+TENSORBOARD_LOGS_PATH=""#<Specify path>
+VOCAB_FILE="" #<Specify path to file>/gpt2-vocab.json
+MERGE_FILE="" #<Specify path to file>/gpt2-merges.txt
+DATA_PATH="" #<Specify path and file prefix>_text_document
+
+docker run \
+  --gpus=all \
+  --ipc=host \
+  --workdir /workspace/megatron-lm \
+  -v /path/to/data:/path/to/data \
+  -v /path/to/megatron-lm:/workspace/megatron-lm \
+  megatron-lm nvcr.io/nvidia/pytorch:24.01-py3 \
+  bash examples/gpt3/train_gpt3_175b_distributed.sh $CHECKPOINT_PATH $TENSORBOARD_LOGS_PATH $VOCAB_FILE $MERGE_FILE $DATA_PATH "
+
+```
+NOTE: Depending on the environment you are running it the above command might like slightly different.
+
+
+## 2. Configurations
+<a id="markdown-configurations" name="configurations"></a>
+The example in this folder shows you how to run 175B model. There are other configs you could run as well
+
+### 345M
+```
+       --num-layers 12 \
+       --hidden-size 512 \
+       --num-attention-heads 8 \
+       --seq-length 1024 \
+       --tensor-model-parallel-size 1 \
+       --pipeline-model-parallel-size 1 \
+
+```
+
+### 857M
+```
+       --num-layers 24 \
+       --hidden-size 1024 \
+       --num-attention-heads 16 \
+       --seq-length 2048 \
+       --tensor-model-parallel-size 1 \
+       --pipeline-model-parallel-size 1 \
+
+```

examples/gpt3/gpt_config.yaml

+# WARNING: Yaml configs is currently an experimental feature
+language_model:
+  # model architecture
+  num_layers: 24
+  hidden_size: 1024
+  num_attention_heads: 16
+  num_query_groups: null
+
+  ffn_hidden_size: null
+  kv_channels: null
+  hidden_dropout: 0.0
+  attention_dropout: 0.0
+  fp32_residual_connection: False
+
+  apply_residual_connection_post_layernorm: False
+  layernorm_epsilon: 1.e-5
+  layernorm_zero_centered_gamma: True
+  add_bias_linear: False
+  bias_activation_fusion: False
+  add_qkv_bias: False
+  gated_linear_unit: False
+  activation_func: swiglu
+  num_moe_experts: null
+  rotary_interleaved: False
+  window_size: null
+
+  # initialization
+  init_method: null
+  init_method_std: 0.02
+  output_layer_init_method: null
+
+  # mixed-precision
+  apply_query_key_layer_scaling: False
+  attention_softmax_in_fp32: False
+
+  # fusion
+  bias_swiglu_fusion: True
+  masked_softmax_fusion: True
+  persist_layer_norm: False
+  memory_efficient_layer_norm: False
+  bias_dropout_fusion: True
+  apply_rope_fusion: True
+
+  # activation recomputation
+  recompute_granularity: null
+  recompute_method: null
+  recompute_num_layers: null
+  distribute_saved_activations: null
+
+  # fp8 related
+  fp8: null
+  fp8_margin: 0
+  fp8_interval: 1
+  fp8_amax_history_len: 1
+  fp8_amax_compute_algo: "most_recent"
+  fp8_wgrad: True
+
+  # miscellaneous
+  clone_scatter_output_in_embedding: True
+
+  normalization: "LayerNorm"  # alt value supported by TE: "RMSNorm"
+
+  # MoE related
+  moe_router_load_balancing_type: "aux_loss"
+  moe_router_topk: 2
+  moe_grouped_gemm: False
+  moe_aux_loss_coeff: 0  # 1e-2 would be a good start value for load balance loss.
+  moe_z_loss_coeff: null  # 1e-3 would be a good start value for z-loss
+  moe_input_jitter_eps: null
+  moe_token_dropping: False
+
+model_parallel:
+  # Model parallelism
+  tensor_model_parallel_size: 1
+  context_parallel_size: 1
+  pipeline_model_parallel_size: 1
+  virtual_pipeline_model_parallel_size: null
+  sequence_parallel: True
+  expert_model_parallel_size: 1
+
+  # Initialization
+  perform_initialization: True
+  use_cpu_initialization: null
+
+  # Training
+  fp16: False
+  bf16: True
+  params_dtype: null # Set from above arguments for core
+  timers: null
+
+  # Optimizations
+  gradient_accumulation_fusion: True
+  async_tensor_model_parallel_allreduce: True
+  tp_comm_overlap: False
+
+  # Debug Options
+  tp_comm_split_ag: True
+  tp_comm_atomic_ag: True
+  tp_comm_split_rs: True
+  tp_comm_atomic_rs: True
+  tp_comm_bulk_wgrad: True
+  tp_comm_bulk_dgrad: True
+
+  # Parallelism
+  finalize_model_grads_func: null
+
+  # Pipeline Parallel
+  pipeline_dtype: null
+  grad_scale_func: null
+  enable_autocast: False
+  autocast_dtype: null
+  variable_seq_lengths: False
+  num_microbatches_with_partial_activation_checkpoints: null
+  overlap_p2p_comm: False
+  batch_p2p_comm: True
+  batch_p2p_sync: True
+  use_ring_exchange_p2p: False
+  deallocate_pipeline_outputs: False
+  no_sync_func: null
+  grad_sync_func: null
+  param_sync_func: null
+  pipeline_model_parallel_split_rank: null
+
+  # CPU Offloading
+  cpu_offloading: False
+  cpu_offloading_num_layers: 0
+  _cpu_offloading_context: null
+  cpu_offloading_weights: False
+  cpu_offloading_activations: True
+
+  # Timing
+  barrier_with_L1_time: True
+
+# training:
+use_legacy_models: False
+spec: null
+micro_batch_size: 2
+global_batch_size: 128
+rampup_batch_size: [32, 32, 65324160] 
+check_for_nan_in_loss_and_grad: True
+num_layers_per_virtual_pipeline_stage: null
+
+encoder_num_layers: null
+decoder_num_layers: null
+rotary_seq_len_interpolation_factor: null
+add_position_embedding: False
+make_vocab_size_divisible_by: 128
+group_query_attention: False
+
+
+exit_signal_handler: False
+exit_duration_in_mins: null
+exit_interval: null
+
+untie_embeddings_and_output_weights: True
+position_embedding_type: rope
+rotary_percent: 0.5
+openai_gelu: False
+squared_relu: False
+swiglu: True
+onnx_safe: null
+bert_binary_head: True
+max_position_embeddings: 4096
+
+transformer_impl: local
+use_flash_attn: False
+seed: 1234
+data_parallel_random_init: False
+
+# Optimizer
+optimizer: adam
+lr: 2.5e-4
+lr_decay_style: cosine
+lr_decay_iters: null
+lr_decay_samples: 255126953
+lr_warmup_fraction: null
+lr_warmup_iters: 0
+lr_warmup_samples: 81381
+lr_warmup_init: 0.0
+min_lr: 2.5e-5
+weight_decay: 0.1
+start_weight_decay: null
+end_weight_decay: null
+weight_decay_incr_style: constant
+clip_grad: 1.0
+adam_beta1: 0.9
+adam_beta2: 0.95
+adam_eps: 1.e-08
+sgd_momentum: 0.9
+override_opt_param_scheduler: False
+use_checkpoint_opt_param_scheduler: False
+
+# checkpointing arguments
+save: null
+save_interval: 20000
+no_save_optim: null
+no_save_rng: null
+load: null
+no_load_optim: null
+no_load_rng: null
+finetune: False
+use_checkpoint_args: False
+exit_on_missing_checkpoint: False
+
+# loss arguments
+loss_scale: null
+initial_loss_scale: 4294967296
+min_loss_scale: 1.0
+loss_scale_window: 1000 
+hysteresis: 2
+accumulate_allreduce_grads_in_fp32: False
+fp16_lm_cross_entropy: False
+
+# distributed arguments
+distributed_backend: nccl
+distributed_timeout_minutes: 10
+overlap_grad_reduce: False
+align_grad_reduce: True
+overlap_param_gather: False
+align_param_gather: False
+scatter_gather_tensors_in_pipeline: True
+local_rank: null
+lazy_mpu_init: null
+empty_unused_memory_level: 0
+standalone_embedding_stage: False
+use_distributed_optimizer: False
+nccl_communicator_config_path: null
+
+train_iters: null
+eval_iters: 32
+eval_interval: 2000
+skip_train: False
+
+adlr_autoresume: False
+adlr_autoresume_interval: 1000
+
+# garbage collection
+manual_gc: False
+manual_gc_interval: 0
+manual_gc_eval: True
+
+tp_comm_overlap_cfg: null
+
+#data
+data_path: null
+split: '99,1,0'
+train_data_path: null
+valid_data_path: null
+test_data_path: null
+data_cache_path: null
+mock_data: False
+vocab_size: null
+vocab_file: null
+merge_file: null
+vocab_extra_ids: 0
+seq_length: 4096
+encoder_seq_length: null
+decoder_seq_length: null
+retriever_seq_length: 256
+sample_rate: 1.0
+mask_prob: 0.15
+short_seq_prob: 0.1
+num_workers: 2
+tokenizer_type: GPTSentencePieceTokenizer
+tokenizer_model: null
+reset_position_ids: False
+reset_attention_mask: False
+eod_mask_loss: False
+train_samples: 268554688
+dataloader_type: null
+
+#profile:
+profile: False
+profile_ranks: [0]
+profile_step_end: 12
+profile_step_start: 10
+
+#logging:
+log_params_norm: True
+log_num_zeros_in_grad: True
+log_throughput: False
+log_progress: False
+timing_log_level: 0
+timing_log_option: minmax
+tensorboard_log_interval: 1
+tensorboard_queue_size: 1000
+log_timers_to_tensorboard: False
+log_validation_ppl_to_tensorboard: False
+log_memory_to_tensorboard: False
+log_world_size_to_tensorboard: False
+log_loss_scale_to_tensorboard: True
+wandb_project: ''
+wandb_exp_name: ''
+wandb_save_dir: ''
+enable_one_logger: True
+one_logger_project: megatron-lm
+one_logger_run_name: null
+log_interval: 100
+tensorboard_dir: null

examples/gpt3/train_gpt3_175b_distributed.sh

+#!/bin/bash
+
+# Runs the "175B" parameter model
+
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+GPUS_PER_NODE=8
+# Change for multinode config
+MASTER_ADDR=localhost
+MASTER_PORT=6000
+NUM_NODES=1
+NODE_RANK=0
+WORLD_SIZE=$(($GPUS_PER_NODE*$NUM_NODES))
+
+CHECKPOINT_PATH=$1 #<Specify path>
+TENSORBOARD_LOGS_PATH=$2 #<Specify path>
+VOCAB_FILE=$3 #<Specify path to file>/gpt2-vocab.json
+MERGE_FILE=$4 #<Specify path to file>/gpt2-merges.txt
+DATA_PATH=$5 #<Specify path and file prefix>_text_document
+
+DISTRIBUTED_ARGS=(
+    --nproc_per_node $GPUS_PER_NODE 
+    --nnodes $NUM_NODES 
+    --master_addr $MASTER_ADDR 
+    --master_port $MASTER_PORT
+)
+
+GPT_MODEL_ARGS=(
+    --num-layers 96 
+    --hidden-size 12288 
+    --num-attention-heads 96 
+    --seq-length 2048 
+    --max-position-embeddings 2048 
+    --attention-backend auto # Can use (flash/fused/unfused/local)
+)
+
+TRAINING_ARGS=(
+    --micro-batch-size 1 
+    --global-batch-size 1536 
+    --rampup-batch-size 16 16 5859375 
+    --train-iters 500000 
+    --weight-decay 0.1 
+    --adam-beta1 0.9 
+    --adam-beta2 0.95 
+    --init-method-std 0.006 
+    --clip-grad 1.0 
+    --fp16
+    --lr 6.0e-5 
+    --lr-decay-style cosine 
+    --min-lr 6.0e-6
+    --lr-warmup-fraction .001 
+    --lr-decay-iters 430000 
+)
+
+MODEL_PARALLEL_ARGS=(
+	--tensor-model-parallel-size 8 
+	--pipeline-model-parallel-size 16 
+)
+
+DATA_ARGS=(
+    --data-path $DATA_PATH 
+    --vocab-file $VOCAB_FILE 
+    --merge-file $MERGE_FILE 
+    --split 949,50,1
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-interval 100
+    --save-interval 10000 
+    --eval-interval 1000 
+    --save $CHECKPOINT_PATH 
+    --load $CHECKPOINT_PATH 
+    --eval-iters 10
+    --tensorboard-dir $TENSORBOARD_LOGS_PATH 
+)
+
+torchrun ${DISTRIBUTED_ARGS[@]} pretrain_gpt.py \
+    ${GPT_MODEL_ARGS[@]} \
+    ${TRAINING_ARGS[@]} \
+    ${MODEL_PARALLEL_ARGS[@]} \
+    ${DATA_ARGS[@]} \
+    ${EVAL_AND_LOGGING_ARGS[@]}

examples/inference/llama_mistral/huggingface_reference.py

+import argparse
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+
+# Set up argument parsing
+parser = argparse.ArgumentParser(description="Script for text generation with a specific model and prompt.")
+parser.add_argument('--prompt', type=str, required=True, help="Prompt text to use for text generation")
+parser.add_argument('--model-path', type=str, required=True, help="Path to the Huggingface model checkpoint")
+
+# Parse command-line arguments
+args = parser.parse_args()
+
+model_path = args.model_path
+prompt = args.prompt
+
+config = AutoConfig.from_pretrained(model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path, config=config)
+model = AutoModelForCausalLM.from_pretrained(model_path, config=config).cuda()
+
+inputs = tokenizer(prompt, return_tensors="pt")
+for key in inputs:
+    inputs[key] = inputs[key].cuda()
+# top_k, top_p and do_sample are set for greedy argmax based sampling
+
+outputs = model.generate(**inputs, max_length=100, do_sample=False, top_p=0, top_k=0, temperature=1.0)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
\ No newline at end of file