[model] Refactor Qwen3-VL and Ministral3 fine-tuning scripts

examples/models/vlm/ministral3/README.md

@@ -85,9 +85,12 @@ The current date is {today}.
 
 - See: [bridge.recipes.ministral3](../../apidocs/bridge/bridge.recipes.ministral3.md)
 - Available recipes:
-  - `ministral3_3b_finetune_config`: Finetuning for 3B VL model with PEFT support
-  - `ministral3_8b_finetune_config`: Finetuning for 8B VL model with PEFT support
-  - `ministral3_14b_finetune_config`: Finetuning for 14B VL model with PEFT support
+  - `ministral3_3b_sft_config`: Finetuning for 3B VL model
+  - `ministral3_8b_sft_config`: Finetuning for 8B VL model
+  - `ministral3_14b_sft_config`: Finetuning for 14B VL model
+  - `ministral3_3b_peft_config`: Finetuning for 3B VL model with PEFT support
+  - `ministral3_8b_peft_config`: Finetuning for 8B VL model with PEFT support
+  - `ministral3_14b_peft_config`: Finetuning for 14B VL model with PEFT support
 
 Before training, ensure the following environment variables are set:
 1. `SAVE_DIR`: checkpoint and log saving directory
@@ -101,15 +104,11 @@ Pretraining is not verified for this model.
 
 ### Supervised Fine-Tuning (SFT)
 
-See the [sft.sh](sft.sh) script for full parameter fine-tuning with configurable model parallelisms.
-
-W&B report coming soon.
+See the [sft_unpacked.sh](sft_unpacked.sh) script for full parameter fine-tuning with configurable model parallelisms.
 
 ### Parameter-Efficient Fine-Tuning (PEFT) with LoRA
 
-See the [peft.sh](peft.sh) script for LoRA fine-tuning with configurable tensor and pipeline parallelism.
-
-W&B report coming soon.
+See the [peft_unpacked.sh](peft_unpacked.sh) script for LoRA fine-tuning with configurable tensor and pipeline parallelism.
 
 ### Recommended Configurations
 
@@ -124,6 +123,9 @@ W&B report coming soon.
 
 **Note:** LoRA/DoRA significantly reduces memory requirements, allowing for larger batch sizes and fewer GPUs.
 
+### Expected Training Dynamics
+We provide a [Weights & Biases report](https://api.wandb.ai/links/nvidia-nemo-fw-public/h32cflfn) for the expected loss curves and grad norms.
+
 ## Evaluation
 
 Coming soon.

examples/models/vlm/ministral3/peft.sh → examples/models/vlm/ministral3/peft_unpacked.sh

@@ -29,12 +29,12 @@ PRETRAINED_CHECKPOINT=${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16
 MODEL_NAME=ministral3_3b
 DATASET_NAME=cord_v2
 SEQ_LENGTH=4096
-TRAIN_ITERS=50
-GLOBAL_BATCH_SIZE=32
+TRAIN_ITERS=100
+GLOBAL_BATCH_SIZE=16
 MICRO_BATCH_SIZE=1
-EVAL_ITERS=10
-LR=0.0002
-MIN_LR=0.00002
+EVAL_ITERS=20
+LR=0.00005
+MIN_LR=0.000005
 LR_WARMUP_ITERS=10
 LOG_INTERVAL=1
 WANDB_PROJECT=megatron-bridge-${DATASET_NAME}
@@ -47,15 +47,15 @@ for config in "${PARALLELISM_CONFIGS[@]}"; do
 
     echo "Running LoRA finetuning with TP=$TP, PP=$PP"
     uv run --no-sync python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
-        --recipe ${MODEL_NAME}_finetune_config \
+        --recipe ${MODEL_NAME}_peft_config \
         --step_func vlm_step \
         --peft_scheme lora \
         checkpoint.pretrained_checkpoint=$PRETRAINED_CHECKPOINT \
         model.seq_length=$SEQ_LENGTH \
         train.train_iters=$TRAIN_ITERS \
         train.global_batch_size=$GLOBAL_BATCH_SIZE \
         train.micro_batch_size=$MICRO_BATCH_SIZE \
-        train.eval_iters=$EVAL_ITERS \
+        validation.eval_iters=$EVAL_ITERS \
         optimizer.lr=$LR \
         optimizer.min_lr=$MIN_LR \
         scheduler.lr_warmup_iters=$LR_WARMUP_ITERS \
@@ -65,6 +65,7 @@ for config in "${PARALLELISM_CONFIGS[@]}"; do
         logger.wandb_exp_name=${MODEL_NAME}_${DATASET_NAME}_lora_tp${TP}_pp${PP} \
         dataset.maker_name=make_${DATASET_NAME}_dataset \
         dataset.seq_length=$SEQ_LENGTH \
+        dataset.pack_sequences_in_batch=False \
         model.tensor_model_parallel_size=$TP \
         model.pipeline_model_parallel_size=$PP
 done

examples/models/vlm/ministral3/sft.sh → examples/models/vlm/ministral3/sft_unpacked.sh

@@ -29,12 +29,12 @@ PRETRAINED_CHECKPOINT=${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16
 MODEL_NAME=ministral3_3b
 DATASET_NAME=cord_v2
 SEQ_LENGTH=4096
-TRAIN_ITERS=50
-GLOBAL_BATCH_SIZE=32
+TRAIN_ITERS=100
+GLOBAL_BATCH_SIZE=16
 MICRO_BATCH_SIZE=1
-EVAL_ITERS=10
-LR=0.00005
-MIN_LR=0.000005
+EVAL_ITERS=20
+LR=0.00001
+MIN_LR=0.000001
 LR_WARMUP_ITERS=10
 LOG_INTERVAL=1
 WANDB_PROJECT=megatron-bridge-${DATASET_NAME}
@@ -47,14 +47,14 @@ for config in "${PARALLELISM_CONFIGS[@]}"; do
 
     echo "Running full finetuning with TP=$TP, PP=$PP"
     uv run --no-sync python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
-        --recipe ${MODEL_NAME}_finetune_config \
+        --recipe ${MODEL_NAME}_sft_config \
         --step_func vlm_step \
         checkpoint.pretrained_checkpoint=$PRETRAINED_CHECKPOINT \
         model.seq_length=$SEQ_LENGTH \
         train.train_iters=$TRAIN_ITERS \
         train.global_batch_size=$GLOBAL_BATCH_SIZE \
         train.micro_batch_size=$MICRO_BATCH_SIZE \
-        train.eval_iters=$EVAL_ITERS \
+        validation.eval_iters=$EVAL_ITERS \
         optimizer.lr=$LR \
         optimizer.min_lr=$MIN_LR \
         scheduler.lr_warmup_iters=$LR_WARMUP_ITERS \
@@ -64,6 +64,7 @@ for config in "${PARALLELISM_CONFIGS[@]}"; do
         logger.wandb_exp_name=${MODEL_NAME}_${DATASET_NAME}_sft_tp${TP}_pp${PP} \
         dataset.maker_name=make_${DATASET_NAME}_dataset \
         dataset.seq_length=$SEQ_LENGTH \
+        dataset.pack_sequences_in_batch=False \
         model.tensor_model_parallel_size=$TP \
         model.pipeline_model_parallel_size=$PP
 done

examples/models/vlm/qwen3_vl/README.md

@@ -106,15 +106,11 @@ Before training, ensure the following environment variables are set:
 See the [sft_unpacked.sh](sft_unpacked.sh) script for full parameter fine-tuning with configurable model parallelisms, with unpacked sequences.
 See the [sft.sh](sft.sh) script for full parameter fine-tuning with sequence-packing.
 
-W&B report coming soon.
-
 ### Parameter-Efficient Fine-Tuning (PEFT) with LoRA
 
 See the [peft_unpacked.sh](peft_unpacked.sh) script for LoRA fine-tuning with configurable tensor and pipeline parallelism, with unpacked sequences.
 See the [peft.sh](peft.sh) script for LoRA fine-tuning with sequence-packing.
 
-W&B report coming soon.
-
 **Note:** LoRA/DoRA significantly reduces memory requirements, allowing for larger batch sizes and fewer GPUs.
 
 ## Finetuning with Energon Dataset
@@ -129,7 +125,11 @@ field_map:
   conversation: json
 ```
 
-Then, update the dataset path (`dataset.path=/path/to/energon/dataset`) in [sft_energon.sh](sft_energon.sh) and run the script.
+Then, update the dataset path (`dataset.path=/path/to/energon/dataset`) in [peft_energon.sh](peft_energon.sh) and run the script.
+
+
+### Expected Training Dynamics
+We provide a [Weights & Biases report](https://api.wandb.ai/links/nvidia-nemo-fw-public/lczz4ixx) for the expected loss curves and grad norms.
 
 ## Evaluation
 

examples/models/vlm/qwen3_vl/peft.sh

@@ -25,10 +25,11 @@ PRETRAINED_CHECKPOINT=${WORKSPACE}/models/Qwen3-VL-8B-Instruct
 MODEL_NAME=qwen3_vl_8b
 DATASET_NAME=cord_v2
 SEQ_LENGTH=4096
-TRAIN_ITERS=50
-GLOBAL_BATCH_SIZE=32
+TRAIN_ITERS=100
+GLOBAL_BATCH_SIZE=16
 MICRO_BATCH_SIZE=2
-EVAL_ITERS=10
+EVAL_ITERS=20
+EVAL_INTERVAL=20
 LR=0.00005
 MIN_LR=0.000005
 LR_WARMUP_ITERS=10
@@ -45,15 +46,16 @@ for pack_config in "${SEQ_PACKING_CONFIGS[@]}"; do
         IFS=',' read -r EP TP PP CP <<< "$par_config"
         echo "Running LoRA finetuning pack_sequences_in_batch=$pack_config with EP=$EP TP=$TP PP=$PP CP=$CP"
         uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
-            --recipe ${MODEL_NAME}_finetune_config \
+            --recipe ${MODEL_NAME}_peft_config \
             --step_func qwen3_vl_step \
             --peft_scheme lora \
             checkpoint.pretrained_checkpoint=$PRETRAINED_CHECKPOINT \
             model.seq_length=$SEQ_LENGTH \
             train.train_iters=$TRAIN_ITERS \
             train.global_batch_size=$GLOBAL_BATCH_SIZE \
             train.micro_batch_size=$MICRO_BATCH_SIZE \
-            train.eval_iters=$EVAL_ITERS \
+            validation.eval_iters=$EVAL_ITERS \
+            validation.eval_interval=$EVAL_INTERVAL \
             optimizer.lr=$LR \
             optimizer.min_lr=$MIN_LR \
             scheduler.lr_warmup_iters=$LR_WARMUP_ITERS \
@@ -80,10 +82,11 @@ PRETRAINED_CHECKPOINT=${WORKSPACE}/models/Qwen3-VL-30B-A3B-Instruct
 MODEL_NAME=qwen3_vl_30b_a3b
 DATASET_NAME=cord_v2
 SEQ_LENGTH=4096
-TRAIN_ITERS=50
-GLOBAL_BATCH_SIZE=32
+TRAIN_ITERS=100
+GLOBAL_BATCH_SIZE=16
 MICRO_BATCH_SIZE=2
-EVAL_ITERS=10
+EVAL_ITERS=20
+EVAL_INTERVAL=20
 LR=0.00005
 MIN_LR=0.000005
 LR_WARMUP_ITERS=10
@@ -100,15 +103,16 @@ for pack_config in "${SEQ_PACKING_CONFIGS[@]}"; do
         IFS=',' read -r EP TP PP CP <<< "$par_config"
         echo "Running LoRA finetuning pack_sequences_in_batch=$pack_config with EP=$EP TP=$TP PP=$PP CP=$CP"
         uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
-            --recipe ${MODEL_NAME}_finetune_config \
+            --recipe ${MODEL_NAME}_peft_config \
             --step_func qwen3_vl_step \
             --peft_scheme lora \
             checkpoint.pretrained_checkpoint=$PRETRAINED_CHECKPOINT \
             model.seq_length=$SEQ_LENGTH \
             train.train_iters=$TRAIN_ITERS \
             train.global_batch_size=$GLOBAL_BATCH_SIZE \
             train.micro_batch_size=$MICRO_BATCH_SIZE \
-            train.eval_iters=$EVAL_ITERS \
+            validation.eval_iters=$EVAL_ITERS \
+            validation.eval_interval=$EVAL_INTERVAL \
             optimizer.lr=$LR \
             optimizer.min_lr=$MIN_LR \
             scheduler.lr_warmup_iters=$LR_WARMUP_ITERS \

examples/models/vlm/qwen3_vl/sft_energon.sh → examples/models/vlm/qwen3_vl/peft_energon.sh

@@ -25,10 +25,11 @@ PRETRAINED_CHECKPOINT=${WORKSPACE}/models/Qwen3-VL-8B-Instruct
 MODEL_NAME=qwen3_vl_8b
 DATASET_NAME=energon
 SEQ_LENGTH=4096
-TRAIN_ITERS=50
-GLOBAL_BATCH_SIZE=32
+TRAIN_ITERS=100
+GLOBAL_BATCH_SIZE=16
 MICRO_BATCH_SIZE=2
-EVAL_ITERS=10
+EVAL_ITERS=20
+EVAL_INTERVAL=20
 LR=0.00005
 MIN_LR=0.000005
 LR_WARMUP_ITERS=10
@@ -47,20 +48,20 @@ for pack_config in "${SEQ_PACKING_CONFIGS[@]}"; do
         IFS=',' read -r EP TP PP CP N_PROC <<< "$par_config"
         echo "Running LoRA finetuning pack_sequences_in_batch=$pack_config with EP=$EP TP=$TP PP=$PP CP=$CP N_PROC=$N_PROC"
         uv run python -m torch.distributed.run --nproc_per_node=$N_PROC scripts/training/run_recipe.py \
-            --recipe ${MODEL_NAME}_finetune_config \
+            --recipe ${MODEL_NAME}_peft_energon_config \
             --step_func qwen3_vl_step \
             --peft_scheme lora \
-            --dataset_type energon \
             checkpoint.pretrained_checkpoint=$PRETRAINED_CHECKPOINT \
             model.seq_length=$SEQ_LENGTH \
             train.train_iters=$TRAIN_ITERS \
             train.global_batch_size=$GLOBAL_BATCH_SIZE \
             train.micro_batch_size=$MICRO_BATCH_SIZE \
-            train.eval_iters=$EVAL_ITERS \
+            validation.eval_iters=$EVAL_ITERS \
+            validation.eval_interval=$EVAL_INTERVAL \
             optimizer.lr=$LR \
             optimizer.min_lr=$MIN_LR \
             scheduler.lr_warmup_iters=$LR_WARMUP_ITERS \
-            checkpoint.save=${WORKSPACE}/results/${MODEL_NAME}_lora_seq_pack_${pack_config}_cp${CP} \
+            checkpoint.save=${WORKSPACE}/results/${MODEL_NAME}_energon_lora_seq_pack_${pack_config}_cp${CP} \
             logger.log_interval=$LOG_INTERVAL \
             logger.wandb_project=$WANDB_PROJECT \
             logger.wandb_exp_name=${MODEL_NAME}_${DATASET_NAME}_lora_seq_pack_${pack_config}_cp${CP} \

examples/models/vlm/qwen3_vl/peft_unpacked.sh

@@ -25,33 +25,35 @@ PRETRAINED_CHECKPOINT=${WORKSPACE}/models/Qwen3-VL-8B-Instruct
 MODEL_NAME=qwen3_vl_8b
 DATASET_NAME=cord_v2
 SEQ_LENGTH=4096
-TRAIN_ITERS=50
-GLOBAL_BATCH_SIZE=32
-MICRO_BATCH_SIZE=1
-EVAL_ITERS=10
+TRAIN_ITERS=100
+GLOBAL_BATCH_SIZE=16
+MICRO_BATCH_SIZE=2
+EVAL_ITERS=20
+EVAL_INTERVAL=20
 LR=0.00005
 MIN_LR=0.000005
 LR_WARMUP_ITERS=10
 LOG_INTERVAL=1
 WANDB_PROJECT=megatron-bridge-${DATASET_NAME}
 
 # TP/PP combinations: "TP,PP"
-PARALLELISM_CONFIGS=("2,1" "1,2")
+PARALLELISM_CONFIGS=("4,1" "2,1")
 
 for config in "${PARALLELISM_CONFIGS[@]}"; do
     IFS=',' read -r TP PP <<< "$config"
 
     echo "Running LoRA finetuning with TP=$TP, PP=$PP"
-    uv run python -m torch.distributed.run --nproc_per_node=2 scripts/training/run_recipe.py \
-        --recipe ${MODEL_NAME}_finetune_config \
+    uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
+        --recipe ${MODEL_NAME}_peft_config \
         --step_func qwen3_vl_step \
         --peft_scheme lora \
         checkpoint.pretrained_checkpoint=$PRETRAINED_CHECKPOINT \
         model.seq_length=$SEQ_LENGTH \
         train.train_iters=$TRAIN_ITERS \
         train.global_batch_size=$GLOBAL_BATCH_SIZE \
         train.micro_batch_size=$MICRO_BATCH_SIZE \
-        train.eval_iters=$EVAL_ITERS \
+        validation.eval_iters=$EVAL_ITERS \
+        validation.eval_interval=$EVAL_INTERVAL \
         optimizer.lr=$LR \
         optimizer.min_lr=$MIN_LR \
         scheduler.lr_warmup_iters=$LR_WARMUP_ITERS \
@@ -71,33 +73,35 @@ PRETRAINED_CHECKPOINT=${WORKSPACE}/models/Qwen3-VL-30B-A3B-Instruct
 MODEL_NAME=qwen3_vl_30b_a3b
 DATASET_NAME=cord_v2
 SEQ_LENGTH=4096
-TRAIN_ITERS=50
-GLOBAL_BATCH_SIZE=32
-MICRO_BATCH_SIZE=1
-EVAL_ITERS=10
+TRAIN_ITERS=100
+GLOBAL_BATCH_SIZE=16
+MICRO_BATCH_SIZE=2
+EVAL_ITERS=20
+EVAL_INTERVAL=20
 LR=0.00005
 MIN_LR=0.000005
 LR_WARMUP_ITERS=10
 LOG_INTERVAL=1
 WANDB_PROJECT=megatron-bridge-${DATASET_NAME}
 
 # EP/TP/PP combinations: "EP,TP,PP" configurations
-PARALLELISM_CONFIGS=("8,1,1" "4,1,1" "2,1,1")
+PARALLELISM_CONFIGS=("8,1,1" "4,1,1")
 
 for config in "${PARALLELISM_CONFIGS[@]}"; do
     IFS=',' read -r EP TP PP <<< "$config"
 
     echo "Running LoRA finetuning with EP=$EP, TP=$TP, PP=$PP"
     uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
-        --recipe ${MODEL_NAME}_finetune_config \
+        --recipe ${MODEL_NAME}_peft_config \
         --step_func qwen3_vl_step \
         --peft_scheme lora \
         checkpoint.pretrained_checkpoint=$PRETRAINED_CHECKPOINT \
         model.seq_length=$SEQ_LENGTH \
         train.train_iters=$TRAIN_ITERS \
         train.global_batch_size=$GLOBAL_BATCH_SIZE \
         train.micro_batch_size=$MICRO_BATCH_SIZE \
-        train.eval_iters=$EVAL_ITERS \
+        validation.eval_iters=$EVAL_ITERS \
+        validation.eval_interval=$EVAL_INTERVAL \
         optimizer.lr=$LR \
         optimizer.min_lr=$MIN_LR \
         scheduler.lr_warmup_iters=$LR_WARMUP_ITERS \