diff --git a/examples/deepseekv3.1-671b-torchtitan/convert/config.py b/examples/deepseekv3.1-671b-torchtitan/convert/config.py
new file mode 100644
index 0000000..8e5c8e3
--- /dev/null
+++ b/examples/deepseekv3.1-671b-torchtitan/convert/config.py
@@ -0,0 +1,49 @@
+from truss.base import truss_config
+from truss_train import definitions
+
+BASE_IMAGE = "ghcr.io/pytorch/pytorch-nightly:71739c4-cu12.9.1"
+PROJECT_NAME = "DeepSeek-V3.1-LoRA"
+
+ACCELERATOR = truss_config.Accelerator.H200
+NGPU = 8
+NODE_COUNT = 1
+
+training_runtime = definitions.Runtime(
+    start_commands=[
+        "chmod +x ./run.sh && ./run.sh",
+    ],
+    environment_variables={
+        # "HF_TOKEN": definitions.SecretReference(name="hf_token"),
+        "HF_HUB_ENABLE_HF_TRANSFER": "true",
+        "WANDB_API_KEY": definitions.SecretReference(name="wandb_api_key"),
+        "NGPU": str(NGPU),
+        "NODE_COUNT": str(NODE_COUNT),
+        "LOG_RANK": "0",
+        "HF_TOKEN": definitions.SecretReference(name="hf_access_token"),
+        "GITHUB_TOKEN": definitions.SecretReference(name="github_token"),
+    },
+    cache_config=definitions.CacheConfig(
+        enabled=True,
+        require_cache_affinity=True,
+    ),
+    checkpointing_config=definitions.CheckpointingConfig(
+        enabled=True,
+    )
+)
+
+training_compute = definitions.Compute(
+    accelerator=truss_config.AcceleratorSpec(
+        accelerator=ACCELERATOR,
+        count=NGPU,
+    ),
+    node_count=NODE_COUNT,
+)
+
+training_job = definitions.TrainingJob(
+    image=definitions.Image(base_image=BASE_IMAGE),
+    compute=training_compute,
+    runtime=training_runtime,
+    name="ConvertDistributedCheckpointsIntoHF"
+)
+
+training_project = definitions.TrainingProject(name=PROJECT_NAME, job=training_job)
\ No newline at end of file
diff --git a/examples/deepseekv3.1-671b-torchtitan/convert/run.sh b/examples/deepseekv3.1-671b-torchtitan/convert/run.sh
new file mode 100755
index 0000000..ca3a7f1
--- /dev/null
+++ b/examples/deepseekv3.1-671b-torchtitan/convert/run.sh
@@ -0,0 +1,168 @@
+#!/bin/bash
+set -euo pipefail
+
+# =============================================================================
+# DeepSeek V3 Checkpoint Conversion Script
+# =============================================================================
+# Converts DCP checkpoints to HuggingFace format and copies to BT_CHECKPOINT_DIR
+#
+# This script should be run on the leader node after training completes.
+# It was extracted from run2.sh to allow manual execution when the conversion
+# doesn't complete due to other nodes finishing first.
+#
+# Required environment variables:
+#   BT_TEAM_CACHE_DIR  - Shared cache directory for model weights
+#   BT_CHECKPOINT_DIR  - Directory to copy converted checkpoints to
+#
+# Usage:
+#   ./convert.sh
+# =============================================================================
+
+# =============================================================================
+# Setup torchtitan environment
+# =============================================================================
+apt update
+apt install -y curl git git-lfs
+
+# Configure git credentials if GITHUB_TOKEN is available
+if [ -n "${GITHUB_TOKEN:-}" ]; then
+    echo "Configuring git credentials..."
+    git config --global credential.helper store
+    echo "https://aghilann:${GITHUB_TOKEN}@github.com" > ~/.git-credentials
+    chmod 600 ~/.git-credentials
+    git config --global url."https://aghilann:${GITHUB_TOKEN}@github.com/".insteadOf "https://github.com/"
+fi
+
+curl -LsSf https://astral.sh/uv/install.sh | sh
+source $HOME/.local/bin/env
+
+mkdir -p /workspace/aghilan-workspace
+cd /workspace/aghilan-workspace
+
+if [[ ! -d "torchtitan" ]]; then
+    echo "Cloning torchtitan..."
+    git clone https://github.com/aghilann/torchtitan
+fi
+
+cd torchtitan
+git checkout lora-stuff
+
+if [[ ! -d ".venv" ]]; then
+    echo "Creating virtual environment..."
+    uv venv
+fi
+
+echo "Activating virtual environment..."
+source .venv/bin/activate
+uv sync
+uv pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu128
+uv pip install wandb safetensors
+
+# =============================================================================
+# Configuration
+# =============================================================================
+
+# Required environment variables
+BT_TEAM_CACHE_DIR="${BT_TEAM_CACHE_DIR:?ERROR: BT_TEAM_CACHE_DIR must be set}"
+BT_CHECKPOINT_DIR="${BT_CHECKPOINT_DIR:?ERROR: BT_CHECKPOINT_DIR must be set}"
+
+# Paths
+HF_ASSETS_PATH="${HF_ASSETS_PATH:-${BT_TEAM_CACHE_DIR}/DeepSeek-V3.1-Base}"
+CHECKPOINT_DIR="${CHECKPOINT_DIR:-${BT_TEAM_CACHE_DIR}/hf_checkpoints/dsv3-671b-lora}"
+
+echo "=============================================="
+echo "DeepSeek V3 Checkpoint Conversion"
+echo "=============================================="
+echo ""
+echo "Configuration:"
+echo "  BT_TEAM_CACHE_DIR: ${BT_TEAM_CACHE_DIR}"
+echo "  BT_CHECKPOINT_DIR: ${BT_CHECKPOINT_DIR}"
+echo "  HF_ASSETS_PATH:    ${HF_ASSETS_PATH}"
+echo "  CHECKPOINT_DIR:    ${CHECKPOINT_DIR}"
+echo "=============================================="
+echo ""
+
+# =============================================================================
+# Convert Checkpoints to HuggingFace Format
+# =============================================================================
+echo "=============================================="
+echo "Converting Checkpoints to HuggingFace Format"
+echo "=============================================="
+
+CHECKPOINT_SUBDIR="${CHECKPOINT_DIR}/checkpoint"
+OUTPUT_BASE="${BT_TEAM_CACHE_DIR}/converted_hf_checkpoint"
+
+if [[ -d "${CHECKPOINT_SUBDIR}" ]]; then
+    echo "Found checkpoint directories:"
+    for dir in "${CHECKPOINT_SUBDIR}"/*/; do
+        if [[ -d "$dir" ]]; then
+            dir_name=$(basename "$dir")
+            echo "  - $dir_name"
+        fi
+    done
+    
+    echo ""
+    echo "Starting conversions..."
+    echo ""
+    
+    for dir in "${CHECKPOINT_SUBDIR}"/*/; do
+        if [[ -d "$dir" ]]; then
+            dir_name=$(basename "$dir")
+            input_path="${CHECKPOINT_SUBDIR}/${dir_name}"
+            output_path="${OUTPUT_BASE}_${dir_name}"
+            
+            echo "=========================================="
+            echo "Converting: $dir_name"
+            echo "  Input:  $input_path"
+            echo "  Output: $output_path"
+            echo "=========================================="
+            
+            python scripts/checkpoint_conversion/convert_to_hf.py \
+                "$input_path" \
+                "$output_path" \
+                --model_name deepseek_v3 \
+                --model_flavor 671B_lora \
+                --hf_assets_path "$HF_ASSETS_PATH" \
+                --export_dtype bfloat16 \
+                --adapters-only \
+                --base_model_name_or_path "deepseek-ai/DeepSeek-V3"
+            
+            if [ $? -eq 0 ]; then
+                echo "✓ Successfully converted $dir_name"
+                
+                # Copy converted checkpoint to BT_CHECKPOINT_DIR
+                dest_path="${BT_CHECKPOINT_DIR}/${dir_name}"
+                echo "Copying converted checkpoint to: $dest_path"
+                mkdir -p "${BT_CHECKPOINT_DIR}"
+                cp -r "$output_path" "$dest_path"
+                if [ $? -eq 0 ]; then
+                    echo "✓ Successfully copied to $dest_path"
+                else
+                    echo "✗ Failed to copy to $dest_path"
+                fi
+            else
+                echo "✗ Failed to convert $dir_name"
+            fi
+
+        fi
+    done
+    
+    echo "=============================================="
+    echo "All checkpoint conversions complete!"
+    echo "=============================================="
+    echo ""
+    echo "Converted checkpoints available at:"
+    ls -d "${OUTPUT_BASE}"_* 2>/dev/null || echo "(none found)"
+    echo ""
+    echo "Checkpoints copied to BT_CHECKPOINT_DIR:"
+    ls -la "${BT_CHECKPOINT_DIR}" 2>/dev/null || echo "(none found)"
+else
+    echo "WARNING: No checkpoint subdirectory found at ${CHECKPOINT_SUBDIR}"
+    echo "Skipping conversion."
+    exit 1
+fi
+
+echo ""
+echo "=============================================="
+echo "Conversion complete!"
+echo "=============================================="
diff --git a/examples/deepseekv3.1-671b-torchtitan/load/config.py b/examples/deepseekv3.1-671b-torchtitan/load/config.py
new file mode 100644
index 0000000..1766921
--- /dev/null
+++ b/examples/deepseekv3.1-671b-torchtitan/load/config.py
@@ -0,0 +1,40 @@
+from truss.base import truss_config
+from truss_train import definitions
+
+BASE_IMAGE = "ghcr.io/pytorch/pytorch-nightly:71739c4-cu12.9.1"
+PROJECT_NAME = "DeepSeek-V3.1-LoRA"
+
+ACCELERATOR = truss_config.Accelerator.H200
+NGPU = 1
+NODE_COUNT = 1
+
+training_runtime = definitions.Runtime(
+    start_commands=[
+        "chmod +x ./run.sh && ./run.sh",
+    ],
+    environment_variables={
+        "HF_HUB_ENABLE_HF_TRANSFER": "true",
+        "HF_TOKEN": definitions.SecretReference(name="hf_access_token"),
+    },
+    cache_config=definitions.CacheConfig(
+        enabled=True,
+        require_cache_affinity=True,
+    ),
+)
+
+training_compute = definitions.Compute(
+    accelerator=truss_config.AcceleratorSpec(
+        accelerator=ACCELERATOR,
+        count=NGPU,
+    ),
+    node_count=NODE_COUNT,
+)
+
+training_job = definitions.TrainingJob(
+    image=definitions.Image(base_image=BASE_IMAGE),
+    compute=training_compute,
+    runtime=training_runtime,
+    name="LoadDataIntoVolume"
+)
+
+training_project = definitions.TrainingProject(name=PROJECT_NAME, job=training_job)
diff --git a/examples/deepseekv3.1-671b-torchtitan/load/run.sh b/examples/deepseekv3.1-671b-torchtitan/load/run.sh
new file mode 100644
index 0000000..a350e46
--- /dev/null
+++ b/examples/deepseekv3.1-671b-torchtitan/load/run.sh
@@ -0,0 +1,29 @@
+#!/bin/bash
+set -euo pipefail
+
+# =============================================================================
+# DeepSeek V3.1 Model Loader
+# =============================================================================
+# Downloads DeepSeek-V3.1-Base model to BT_TEAM_CACHE_DIR
+# =============================================================================
+
+# Install dependencies
+pip install huggingface_hub hf_transfer
+
+# Clone torchtitan for the download script
+git clone https://github.com/aghilann/torchtitan
+cd torchtitan
+
+# Download the model assets
+python scripts/download_hf_assets.py \
+    --repo_id deepseek-ai/DeepSeek-V3.1-Base \
+    --assets safetensors config tokenizer \
+    --local_dir $BT_TEAM_CACHE_DIR
+
+echo ""
+echo "=============================================="
+echo "Model downloaded to: $BT_TEAM_CACHE_DIR"
+echo "=============================================="
+echo ""
+echo "Contents:"
+ls -lah $BT_TEAM_CACHE_DIR
diff --git a/examples/deepseekv3.1-671b-torchtitan/train/config.py b/examples/deepseekv3.1-671b-torchtitan/train/config.py
new file mode 100644
index 0000000..decbefd
--- /dev/null
+++ b/examples/deepseekv3.1-671b-torchtitan/train/config.py
@@ -0,0 +1,49 @@
+from truss.base import truss_config
+from truss_train import definitions
+
+BASE_IMAGE = "ghcr.io/pytorch/pytorch-nightly:71739c4-cu12.9.1"
+PROJECT_NAME = "DeepSeek-V3.1-LoRA"
+
+
+ACCELERATOR = truss_config.Accelerator.H200
+NGPU = 8
+NODE_COUNT = 4
+
+training_runtime = definitions.Runtime(
+    start_commands=[
+        "chmod +x ./run.sh && ./run.sh",
+    ],
+    environment_variables={
+        "HF_HUB_ENABLE_HF_TRANSFER": "true",
+        "WANDB_API_KEY": definitions.SecretReference(name="wandb_api_key"),
+        "NGPU": str(NGPU),
+        "NODE_COUNT": str(NODE_COUNT),
+        "LOG_RANK": "0",
+        "HF_TOKEN": definitions.SecretReference(name="hf_access_token"),
+        "GITHUB_TOKEN": definitions.SecretReference(name="github_token"),
+    },
+    cache_config=definitions.CacheConfig(
+        enabled=True,
+        require_cache_affinity=True,
+    ),
+    checkpointing_config=definitions.CheckpointingConfig(
+        enabled=True,
+    )
+)
+
+training_compute = definitions.Compute(
+    accelerator=truss_config.AcceleratorSpec(
+        accelerator=ACCELERATOR,
+        count=NGPU,
+    ),
+    node_count=NODE_COUNT,
+)
+
+training_job = definitions.TrainingJob(
+    image=definitions.Image(base_image=BASE_IMAGE),
+    compute=training_compute,
+    runtime=training_runtime,
+    name="Training"
+)
+
+training_project = definitions.TrainingProject(name=PROJECT_NAME, job=training_job)
\ No newline at end of file
diff --git a/examples/deepseekv3.1-671b-torchtitan/train/run.sh b/examples/deepseekv3.1-671b-torchtitan/train/run.sh
new file mode 100644
index 0000000..3cae939
--- /dev/null
+++ b/examples/deepseekv3.1-671b-torchtitan/train/run.sh
@@ -0,0 +1,177 @@
+#!/bin/bash
+set -euo pipefail
+# Misc utils
+apt update
+apt install -y curl fzf ripgrep git git-lfs tmux htop lsof gh neovim
+apt-get update && apt-get install -y build-essential
+apt-get install nvtop
+
+export TUNNEL_NAME="bt-${BT_TRAINING_JOB_ID}-${BT_NODE_RANK}"
+
+# Configure git credentials if GITHUB_TOKEN is available
+if [ -n "${GITHUB_TOKEN:-}" ]; then
+    echo "Configuring git credentials..."
+    git config --global credential.helper store
+    echo "https://aghilann:${GITHUB_TOKEN}@github.com" > ~/.git-credentials
+    chmod 600 ~/.git-credentials
+    git config --global url."https://aghilann:${GITHUB_TOKEN}@github.com/".insteadOf "https://github.com/"
+fi
+
+curl -LsSf https://astral.sh/uv/install.sh | sh
+
+mkdir aghilan-workspace
+
+cd aghilan-workspace
+echo "Cloning torchtitan"
+git clone https://github.com/basetenlabs/kingkong.git
+cd kingkong
+uv venv
+echo "Creating virtual environment"
+. .venv/bin/activate
+uv sync
+uv pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu128
+uv pip install wandb safetensors
+chmod +x run_train.sh
+
+# =============================================================================
+# DeepSeek V3 671B LoRA Fine-Tuning - Multi-Node Training Script
+# =============================================================================
+# Designed for Baseten multi-node training infrastructure.
+#
+# Required Baseten Environment Variables:
+#   BT_GROUP_SIZE          - Number of nodes in the training group
+#   BT_NUM_GPUS            - Number of GPUs per node
+#   BT_NODE_RANK           - This node's rank (0 to BT_GROUP_SIZE-1)
+#   BT_LEADER_ADDR         - IP address of the leader node (rank 0)
+#   BT_TEAM_CACHE_DIR      - Shared cache directory for model weights
+#   BT_TRAINING_PROJECT_NAME - WandB project name
+#   BT_TRAINING_JOB_NAME   - WandB run/experiment name
+#
+# Usage:
+#   ./scripts/train_deepseek_v3_lora_multinode.sh
+# =============================================================================
+
+# rm -rf /root/.cache/team_artifacts/hf_checkpoints
+
+DEBUG="false"  # Set to 'true' for debug (2-layer model), 'false' for production (671B)
+
+# Configuration - Uses Baseten environment variables
+NNODES="${BT_GROUP_SIZE:?ERROR: BT_GROUP_SIZE must be set}"
+NPROC_PER_NODE="${BT_NUM_GPUS:?ERROR: BT_NUM_GPUS must be set}"
+NODE_RANK="${BT_NODE_RANK:?ERROR: BT_NODE_RANK must be set}"
+MASTER_ADDR="${BT_LEADER_ADDR:?ERROR: BT_LEADER_ADDR must be set}"
+MASTER_PORT="${MASTER_PORT:-29500}"
+
+export NGPU="${BT_NUM_GPUS:?ERROR: BT_NUM_GPUS must be set}"
+
+# Config, assets, and checkpoint paths based on DEBUG flag
+if [[ "${DEBUG}" == "true" ]]; then
+    CONFIG_FILE="${CONFIG_FILE:-./torchtitan/models/deepseek_v3/train_configs/deepseek_aghilora.toml}"
+    HF_ASSETS_PATH="${HF_ASSETS_PATH:-${BT_TEAM_CACHE_DIR}/DeepSeek-v3.1-Base-DEBUG}"
+    CHECKPOINT_DIR="${CHECKPOINT_DIR:-${BT_TEAM_CACHE_DIR}/hf_checkpoints/dsv3-debug-lora}"
+else
+    CONFIG_FILE="${CONFIG_FILE:-./torchtitan/models/deepseek_v3/train_configs/deepseek_v3_671b_lora.toml}"
+    HF_ASSETS_PATH="${HF_ASSETS_PATH:-${BT_TEAM_CACHE_DIR}/DeepSeek-V3.1-Base}"
+    CHECKPOINT_DIR="${CHECKPOINT_DIR:-${BT_TEAM_CACHE_DIR}/hf_checkpoints/dsv3-671b-lora}"
+fi
+
+# WandB settings (from Baseten env)
+WANDB_PROJECT="${BT_TRAINING_PROJECT_NAME:?ERROR: BT_TRAINING_PROJECT_NAME must be set}"
+WANDB_RUN_NAME="${BT_TRAINING_JOB_NAME:?ERROR: BT_TRAINING_JOB_NAME must be set}"
+
+# Environment Setup
+export PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True"
+export NCCL_DEBUG="${NCCL_DEBUG:-WARN}"
+export NCCL_IB_DISABLE="${NCCL_IB_DISABLE:-0}"
+export NCCL_NET_GDR_LEVEL="${NCCL_NET_GDR_LEVEL:-5}"
+export NCCL_TIMEOUT=3600000 # 1 hour
+export WANDB_PROJECT="${WANDB_PROJECT}"
+export WANDB_RUN_NAME="${WANDB_RUN_NAME}"
+
+# Print Configuration
+echo "=============================================="
+echo "DeepSeek V3 LoRA Multi-Node Training"
+echo "=============================================="
+if [[ "${DEBUG}" == "true" ]]; then
+    echo "Mode: DEBUG (2-layer model)"
+else
+    echo "Mode: PRODUCTION (671B)"
+fi
+echo ""
+echo "Baseten Environment:"
+echo "  BT_GROUP_SIZE (nodes):     ${NNODES}"
+echo "  BT_NUM_GPUS (per node):    ${NPROC_PER_NODE}"
+echo "  BT_NODE_RANK:              ${NODE_RANK}"
+echo "  BT_LEADER_ADDR:            ${MASTER_ADDR}"
+echo "  MASTER_PORT:               ${MASTER_PORT}"
+echo ""
+echo "Paths:"
+echo "  CONFIG_FILE:               ${CONFIG_FILE}"
+echo "  HF_ASSETS_PATH:            ${HF_ASSETS_PATH}"
+echo "  CHECKPOINT_DIR:            ${CHECKPOINT_DIR}"
+echo ""
+echo "WandB:"
+echo "  BT_TRAINING_PROJECT_NAME:  ${WANDB_PROJECT}"
+echo "  BT_TRAINING_JOB_NAME:      ${WANDB_RUN_NAME}"
+echo "=============================================="
+
+# Validate Environment
+if [[ ! -f "${CONFIG_FILE}" ]]; then
+    echo "ERROR: Config file not found: ${CONFIG_FILE}"
+    exit 1
+fi
+
+if [[ ! -d "${HF_ASSETS_PATH}" ]]; then
+    echo "ERROR: HF assets not found: ${HF_ASSETS_PATH}"
+    echo "Please download with: python scripts/download_hf_assets.py --repo_id deepseek-ai/DeepSeek-V3.1-Base --local_dir \$BT_TEAM_CACHE_DIR --all"
+    exit 1
+fi
+
+if [[ "${NODE_RANK}" -lt 0 ]] || [[ "${NODE_RANK}" -ge "${NNODES}" ]]; then
+    echo "ERROR: NODE_RANK must be between 0 and $((NNODES - 1))"
+    exit 1
+fi
+
+# Launch Training
+echo ""
+echo "Starting training on node ${NODE_RANK}..."
+echo ""
+
+echo "Checkpoint directory BEFORE training:"
+ls -lah "${CHECKPOINT_DIR}" 2>/dev/null || echo "(directory does not exist yet)"
+echo ""
+
+torchrun \
+    --nnodes="${NNODES}" \
+    --nproc_per_node="${NPROC_PER_NODE}" \
+    --node_rank="${NODE_RANK}" \
+    --rdzv_backend=c10d \
+    --rdzv_endpoint="${MASTER_ADDR}:${MASTER_PORT}" \
+    --local-ranks-filter 0 \
+    --role rank \
+    --tee 3 \
+    -m torchtitan.train \
+    --job.config_file "${CONFIG_FILE}" \
+    --job.dump_folder "${CHECKPOINT_DIR}" \
+    --model.hf_assets_path "${HF_ASSETS_PATH}" \
+    --parallelism.expert_parallel_degree 16 \
+    --parallelism.fsdp_reshard_after_forward always \
+    --checkpoint.export_dtype bfloat16 \
+    --activation_checkpoint.mode full \
+    --training.local_batch_size 2 \
+    --training.steps 1000 \
+    --checkpoint.interval 500 \
+    --optimizer.lr 1.5e-4 \
+    --lr_scheduler.warmup_steps 0 \
+    --lr_scheduler.min_lr_factor 1.0
+
+
+
+echo ""
+echo "Checkpoint directory AFTER training:"
+ls -lah "${CHECKPOINT_DIR}" 2>/dev/null || echo "(directory does not exist)"
+echo ""
+
+echo "=============================================="
+echo "Training completed on node ${NODE_RANK}"
+echo "=============================================="
\ No newline at end of file