同步最新代码

examples/multimodal/nvlm/pretrain_yi_34b_internvit_6b.sh

@@ -75,7 +75,7 @@ OPTIONS=" \
     --decoder-seq-length ${DECODER_SEQ_LEN} \
     --max-position-embeddings ${MAX_POS_EMBED} \
     --tokenizer-type MultimodalTokenizer \
-    --tokenizer-model ${WORKSPACE}/<path to tokenizer> \
+    --tokenizer-model NousResearch/Nous-Hermes-2-Yi-34B \
     --tokenizer-prompt-format nvlm-yi-34b \
     --vocab-size 64000 \
     --make-vocab-size-divisible-by 1 \

examples/multimodal/nvlm/run_text_generation_qwen20_72b_internvit_6b.sh

@@ -97,7 +97,7 @@ do
         --decoder-seq-length ${DECODER_SEQ_LEN} \
         --max-position-embeddings ${MAX_POS_EMBED} \
         --tokenizer-type MultimodalTokenizer \
-        --tokenizer-model <tokenizer model path> \
+        --tokenizer-model Qwen/Qwen2-72B-Instruct \
         --tokenizer-prompt-format qwen2p0 \
         --position-embedding-type rope \
         --rotary-percent 1.0 \

examples/multimodal/nvlm/run_text_generation_qwen25_7b_siglip.sh

+#!/bin/bash
+
+export NCCL_IB_SL=1
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export NVTE_APPLY_QK_LAYER_SCALING=0
+export TOKENIZERS_PARALLELISM="false"
+
+INPUT_IMAGE_PATH="placeholder"
+GROUNDTRUTH_PATH="placeholder"
+
+while [[ $# -gt 0 ]]; do
+    case $1 in
+        -i|--input-image-path)
+            INPUT_IMAGE_PATH="$2"
+            shift
+            shift
+            ;;
+        -o|--output-path)
+            OUTPUT_PATH="$2"
+            shift
+            shift
+            ;;
+        -m|--model-path)
+            MODEL_PATH="$2"
+            shift
+            shift
+            ;;
+        -t|--task)
+            TASK="$2"
+            shift
+            shift
+            ;;
+        -g|--gt-path)
+            GROUNDTRUTH_PATH="$2"
+            shift
+            shift
+            ;;
+        -*|--*)
+            echo "Invalid option $1"
+            exit 1
+            ;;
+    esac
+done
+
+# Please modify these as needed.
+NUM_PARTITIONS=0
+START=0
+END=0
+
+
+SEQ_LEN=256
+DECODER_SEQ_LEN=8192
+EXTRA_ARGS=" --pixel-shuffle --use-tiling --max-num-tiles 12 --use-thumbnail"
+
+for PARTITION_ID in $( eval echo {$START..$END} )
+do
+    torchrun --nproc_per_node 8 examples/multimodal/run_text_generation.py \
+        --attention-softmax-in-fp32 \
+        --transformer-impl transformer_engine \
+        --use-te \
+        --use-checkpoint-args \
+        --normalization RMSNorm \
+        --norm-epsilon 1e-06 \
+        --language-model-type=qwen2.5_7B \
+        --untie-embeddings-and-output-weights \
+        --disable-bias-linear \
+        --position-embedding-type rope \
+        --rotary-percent 1.0 \
+        --rotary-base 1000000 \
+        --swiglu \
+        --attention-dropout 0.0 \
+        --hidden-dropout 0.0 \
+        --tensor-model-parallel-size 4 \
+        --pipeline-model-parallel-size 1 \
+        --group-query-attention \
+        --num-query-groups 4 \
+        --num-layers 28 \
+        --hidden-size 3584 \
+        --ffn-hidden-size 18944 \
+        --add-qkv-bias \
+        --num-attention-heads 28 \
+        --max-position-embeddings 32768  \
+        --no-masked-softmax-fusion \
+        --load ${MODEL_PATH} \
+        --tokenizer-type MultimodalTokenizer \
+        --tokenizer-model Qwen/Qwen2.5-7B-Instruct \
+        --tokenizer-prompt-format qwen2p5 \
+        --bf16 \
+        --micro-batch-size 1 \
+        --seq-length ${SEQ_LEN} \
+        --decoder-seq-length ${DECODER_SEQ_LEN} \
+        --out-seq-length 128 \
+        --temperature 1.0 \
+        --img-h 448 \
+        --img-w 448 \
+        --patch-dim 14 \
+        --seed 153 \
+        --top_k 1 \
+        --no-load-rng \
+        --no-load-optim \
+        --input-image-path ${INPUT_IMAGE_PATH} \
+        --num-partitions ${NUM_PARTITIONS} \
+        --partition-id ${PARTITION_ID} \
+        --output-path ${OUTPUT_PATH} \
+        --gt-path ${GROUNDTRUTH_PATH} \
+        --task ${TASK} \
+        ${EXTRA_ARGS} \
+        --special-tokens "<image>" "<img>" "</img>" \
+        --vision-model-type siglip \
+        --ckpt-format torch
+done

examples/multimodal/nvlm/run_text_generation_yi_34b_internvit_6b.sh

@@ -95,7 +95,7 @@ do
         --decoder-seq-length ${DECODER_SEQ_LEN} \
         --max-position-embeddings ${MAX_POS_EMBED} \
         --tokenizer-type MultimodalTokenizer \
-        --tokenizer-model <tokenizer model path> \
+        --tokenizer-model NousResearch/Nous-Hermes-2-Yi-34B \
         --tokenizer-prompt-format nvlm-yi-34b \
         --vocab-size 64000 \
         --make-vocab-size-divisible-by 1 \
@@ -135,6 +135,6 @@ do
         --gt-path ${GROUNDTRUTH_PATH} \
         ${EXTRA_ARGS} \
         --task ${TASK} \
-        --image-tag-type nlvm \
+        --image-tag-type nvlm \
         --ckpt-format torch
 done

examples/multimodal/nvlm/sft_34b_internvit.sh

@@ -80,7 +80,7 @@ OPTIONS=" \
     --decoder-seq-length ${DECODER_SEQ_LEN} \
     --max-position-embeddings ${MAX_POS_EMBED} \
     --tokenizer-type MultimodalTokenizer \
-    --tokenizer-model ${WORKSPACE}/<tokenizer path> \
+    --tokenizer-model NousResearch/Nous-Hermes-2-Yi-34B \
     --tokenizer-prompt-format nvlm-yi-34b \
     --vocab-size 64000 \
     --make-vocab-size-divisible-by 1 \

examples/multimodal/nvlm/sft_qwen20_72b_internvit_6b.sh

@@ -67,7 +67,7 @@ OPTIONS=" \
     --exit-duration-in-mins 230 \
     --disable-bias-linear \
     --tokenizer-type MultimodalTokenizer \
-    --tokenizer-model ${WORKSPACE}/<tokenizer model path> \
+    --tokenizer-model Qwen/Qwen2-72B-Instruct \
     --tokenizer-prompt-format qwen2p0 \
     --transformer-impl transformer_engine \
     --normalization RMSNorm \

examples/multimodal/pretrain_mistral_clip.sh

@@ -24,11 +24,6 @@ if [[ -z $LOAD_NAME ]]; then
     exit 1
 fi
 
-if [[ -z $TOKENIZER_MODEL ]]; then
-    echo "Please set TOKENIZER_MODEL for tokenizer model name."
-    exit 1
-fi
-
 CHECKPOINT_DIR="${WORKSPACE}/${LOAD_NAME}/checkpoints"
 
 DATA_TRAIN="${SOURCE}/examples/multimodal/pretrain_dataset.yaml"
@@ -93,7 +88,7 @@ OPTIONS=" \
     --eval-iters 10 \
     --eval-interval 1000 \
     --tokenizer-type MultimodalTokenizer \
-    --tokenizer-model ${WORKSPACE}/${TOKENIZER_MODEL} \
+    --tokenizer-model mistralai/Mistral-7B-Instruct-v0.3 \
     --tokenizer-prompt-format mistral \
     --data-path ${DATA_TRAIN} \
     --prompt-path ${SOURCE}/examples/multimodal/manual_prompts.json \

examples/multimodal/run_text_generation.py

@@ -14,11 +14,13 @@ sys.path.append(
 import torch
 import yaml
 from config import EvaluationConfig
-from evaluation_datasets import get_evaluation_dataset
+from evaluation.evaluation_datasets import get_evaluation_dataset
 from model import model_provider
 from multimodal_args import add_multimodal_extra_args
 
 from megatron.core import parallel_state
+from megatron.core.enums import ModelType
+from megatron.core.models.multimodal.llava_model import IMAGE_TOKEN
 from megatron.core.models.vision.clip_vit_model import get_num_image_embeddings
 from megatron.inference.text_generation.api import generate_and_post_process
 from megatron.inference.text_generation.forward_step import ForwardStep
@@ -36,7 +38,7 @@ def add_text_generation_args(parser):
     group.add_argument("--top_p", type=float, default=0.0, help='Top p sampling.')
     group.add_argument("--top_k", type=int, default=0, help='Top k sampling.')
     group.add_argument(
-        "--out-seq-length", type=int, default=1024, help='Length of the output generated text.'
+        "--out-seq-length", type=int, default=128, help='Length of the output generated text.'
     )
     group.add_argument("--output-path", type=str, help='Output file path')
     group.add_argument('--input-image-path', type=str, help="Input image directory")
@@ -206,8 +208,8 @@ def generate_samples(model, config: EvaluationConfig, print_output):
                 if config.task == "VideoMME":
                     output["questions"][0][output_name] = generated
                 else:
-                    output[output_name] = generated
                     output["prompt"] = prompt
+                    output[output_name] = generated
 
                 if config.task == "captioning":
                     output["ground_truth"] = answers
@@ -354,7 +356,7 @@ class VLMForwardStep(ForwardStep):
         )
 
     def __call__(self, tokens, position_ids, attention_mask):
-        num_image_tokens = (tokens == self.model.image_token_index).sum().item()
+        num_image_tokens = (tokens == self.model.module.image_token_index).sum().item()
         num_tokens = tokens.size(1)
         recv_buffer_seq_length = None
         if num_image_tokens > 0:
@@ -406,7 +408,7 @@ def get_conversation(task, question):
             {"role": "system", "content": "Answer the questions."},
             {
                 "role": "user",
-                "content": "<image>\nProvide a one-sentence caption for provided image.",
+                "content": f"{IMAGE_TOKEN}\nProvide a one-sentence caption for provided image.",
             },
         ]
     elif task in ("TextVQA", "VQAv2", "ChartQA"):
@@ -414,13 +416,13 @@ def get_conversation(task, question):
             {"role": "system", "content": "Answer the questions."},
             {
                 "role": "user",
-                "content": f"<image>\n{question}\nAnswer the question using a single word or phrase.",
+                "content": f"{IMAGE_TOKEN}\n{question}\nAnswer the question using a single word or phrase.",
             },
         ]
     elif task in ("OCRBench", "MathVista", "AI2D"):
         conversation = [
             {"role": "system", "content": "Answer the questions."},
-            {"role": "user", "content": f"<image>\n{question}"},
+            {"role": "user", "content": f"{IMAGE_TOKEN}\n{question}"},
         ]
     elif task == "MMMU":
         conversation = [
@@ -441,7 +443,7 @@ def get_conversation(task, question):
 
         conversation = [
             {"role": "system", "content": "Answer the questions."},
-            {"role": "user", "content": f"<image>\n{question}"},
+            {"role": "user", "content": f"{IMAGE_TOKEN}\n{question}"},
         ]
 
     return conversation
@@ -464,11 +466,13 @@ def get_prompt_and_generated(prompt_and_generation, prompt_format):
         prompt = splitted[0]
         generated = splitted[1]
         generated = generated.split("<|im_end|>")[0]
-    elif prompt_format in ("nvlm-yi-34b", "qwen2p0"):
+    elif prompt_format in ("nvlm-yi-34b", "qwen2p0", "qwen2p5"):
         splitted = prompt_and_generation.split("<|im_start|>assistant\n")
         prompt = splitted[0]
         generated = splitted[1]
         generated = generated.split("<|im_end|>")[0]
+    else:
+        raise ValueError(f"Prompt format {prompt_format} is not supported.")
 
     # Remove possible garbage.
     generated = generated.strip()
@@ -489,11 +493,11 @@ def main():
 
     args = get_args()
 
-    def wrapped_model_provider(pre_process, post_process):
-        return model_provider(pre_process, post_process, parallel_output=False)
+    def wrapped_model_provider(pre_process, post_process, add_encoder, add_decoder):
+        return model_provider(pre_process, post_process, add_encoder, add_decoder, parallel_output=False)
 
     # Set up model and load checkpoint.
-    model = get_model(wrapped_model_provider, wrap_with_ddp=False)
+    model = get_model(wrapped_model_provider, model_type=ModelType.encoder_and_decoder, wrap_with_ddp=False)
 
     if args.load is not None:
         _ = load_checkpoint(model, None, None)

examples/multimodal/sft_mistral_clip.sh

@@ -29,11 +29,6 @@ if [[ -z $LOAD_ITER ]]; then
     exit 1
 fi
 
-if [[ -z $TOKENIZER_MODEL ]]; then
-    echo "Please set TOKENIZER_MODEL for tokenizer model name."
-    exit 1
-fi
-
 CHECKPOINT_DIR="${WORKSPACE}/${LOAD_NAME}/checkpoints"
 
 DATA_TRAIN="${SOURCE}/examples/multimodal/sft_dataset.yaml"
@@ -98,7 +93,7 @@ OPTIONS=" \
     --eval-iters 10 \
     --eval-interval 500 \
     --tokenizer-type MultimodalTokenizer \
-    --tokenizer-model ${WORKSPACE}/${TOKENIZER_MODEL} \
+    --tokenizer-model mistralai/Mistral-7B-Instruct-v0.3 \
     --tokenizer-prompt-format mistral \
     --data-path ${DATA_TRAIN} \
     --prompt-path ${SOURCE}/examples/multimodal/manual_prompts.json \

examples/multimodal/text_generation_mistral_clip.sh

@@ -4,12 +4,13 @@ export NCCL_IB_SL=1
 export CUDA_DEVICE_MAX_CONNECTIONS=1
 export NVTE_APPLY_QK_LAYER_SCALING=0
 
+INPUT_IMAGE_PATH="placeholder"
 GROUNDTRUTH_PATH="placeholder"
 NUM_FRAMES=1
 
 while [[ $# -gt 0 ]]; do
     case $1 in
-        --input-image-path)
+        -i|--input-image-path)
             INPUT_IMAGE_PATH="$2"
             shift
             shift
@@ -19,11 +20,6 @@ while [[ $# -gt 0 ]]; do
             shift
             shift
             ;;
-        -g|--groundtruth-path)
-            GROUNDTRUTH_PATH="$2"
-            shift
-            shift
-            ;;
         -o|--output-path)
             OUTPUT_PATH="$2"
             shift
@@ -34,12 +30,7 @@ while [[ $# -gt 0 ]]; do
             shift
             shift
             ;;
-        -t|--tokenizer-path)
-            TOKENIZER_PATH="$2"
-            shift
-            shift
-            ;;
-        --task)
+        -t|--task)
             TASK="$2"
             shift
             shift
@@ -92,7 +83,7 @@ do
         --no-masked-softmax-fusion \
         --load ${MODEL_PATH} \
         --tokenizer-type MultimodalTokenizer \
-        --tokenizer-model ${TOKENIZER_PATH} \
+        --tokenizer-model mistralai/Mistral-7B-Instruct-v0.3 \
         --tokenizer-prompt-format mistral \
         --bf16 \
         --micro-batch-size 1 \

examples/multimodal/train.py

@@ -48,7 +48,7 @@ def get_batch(data_iterator):
     pp_size = get_pipeline_model_parallel_world_size()
     if not is_first_or_last_stage(pp_size, args.encoder_pipeline_model_parallel_size):
         # Note these are all set to None above.
-        return tokens, labels, loss_mask, attention_mask, position_ids, imgs, num_tiles
+        return tokens, labels, loss_mask, attention_mask, position_ids, imgs, num_tiles, packed_seq_params
 
     # Broadcast data.
     torch.cuda.nvtx.range_push("get_data")
@@ -66,7 +66,7 @@ def get_batch(data_iterator):
     cu_lengths = tensor_parallel.broadcast_data(["cu_lengths"], data, torch.int32)["cu_lengths"]
     max_lengths = tensor_parallel.broadcast_data(["max_lengths"], data, torch.int32)["max_lengths"]
 
-    # Dummy image, no image.
+    # No image input (text-only sample) if the dataloader produced a dummy image.
     if imgs.shape == torch.Size([1, 1]):
         # FIXME: text-only data can cause a hang if the vision model is own its own pipeline rank and --freeze-ViT is enabled.
         imgs = torch.tensor([], dtype=torch.float32, device=data_text.device)