Add transformers backend (Dense model only)

[3outeille]

Context

This PR enables:

• Llama-like HF models to work with 4D parallelism: FSDP, CP, TP, PP (and the combinations between them). The following models were tested:
  • meta-llama/Llama-3.2-1B
  • microsoft/phi-2
  • Qwen/Qwen2.5-7B
  • mistralai/Mistral-7B-v0.1
  • ByteDance-Seed/Seed-Coder-8B-Instruct
  • Qwen/Qwen3-4B-Instruct-2507
  • arcee-ai/AFM-4.5B
  • ibm-granite/granite-3b-code-base-2k
  • baidu/ERNIE-4.5-0.3B-Base-PT
  • kyutai/helium-1-preview-2b
  • allenai/OLMo-7B-hf
  • mistralai/Ministral-8B-Instruct-2410
• Patching HF models weights initialisation. Without this, the the loss and grad_norm starts very high

Usage

• Config: torchtitan/torchtitan/experiments/transformers_backend/configs/qwen3_fsdp2_tp2_pp2.toml

...
[model]
- name = "llama3"
+ name = "Qwen/Qwen3-4B-Instruct-2507" 
flavor = "debugmodel"
hf_assets_path = "./tests/assets/tokenizer"
...

• Train: LOG_RANK=7 ./torchtitan/torchtitan/experiments/transformers_backend/run_train.sh

Testing methodology

• Following the converging.md guidelines, I am comparing the baseline FSDP=2 vs FSDP=2 & <other //-ism>
• More precisely, the test_hf_integration.pyis going to do:

    results/
        |_ meta-llama
            |_ Llama-3.2-1B
                |_ debugmodel/
                    |_ seed_checkpoint/
                        |_ config.toml
                        |_ seed.slurm
                        |_ step-0/
                           |_ ....
                    |_ fsdp2_tp1_cp1_pp1/
                        |_ config.toml
                        |_ nd_parallelism.slurm
                        |_ nd_parallelism.log
                    |_ fsdp2_tp2_cp1_pp1/
                        |_ config.toml
                        |_ nd_parallelism.slurm
                        |_ nd_parallelism.log
                        |_ diff_baseline_vs_nd_parallelism.log
                    |_ fsdp2_tp1_cp1_pp2/
                        |_ config.toml
                        |_ nd_parallelism.slurm
                        |_ nd_parallelism.log
                        |_ diff_baseline_vs_nd_parallelism.log
                    |_ fsdp2_tp1_cp2_pp1/
                        |_ config.toml
                        |_ nd_parallelism.slurm
                        |_ nd_parallelism.log
                        |_ diff_baseline_vs_nd_parallelism.log
                    |_ fsdp2_tp1_cp2_pp2/
                        |_ config.toml
                        |_ nd_parallelism.slurm
                        |_ nd_parallelism.log
                        |_ diff_baseline_vs_nd_parallelism.log`
                |_ full/
                ...

• Here is the grid search to test the HF modelling

#!/usr/bin/bash
model_names=(
     "meta-llama/Llama-3.2-1B"
     "microsoft/phi-2" 
     "Qwen/Qwen2.5-7B"
     "mistralai/Mistral-7B-v0.1"
     "ByteDance-Seed/Seed-Coder-8B-Instruct"
     "Qwen/Qwen3-4B-Instruct-2507" 
     "arcee-ai/AFM-4.5B" 
     "ibm-granite/granite-3b-code-base-2k" 
     "baidu/ERNIE-4.5-0.3B-Base-PT" 
     "kyutai/helium-1-preview-2b" 
     "allenai/OLMo-7B-hf"
     "mistralai/Ministral-8B-Instruct-2410" 
)

for model_name in "${model_names[@]}"; do
    rm -rf slurm_results/${model_name}

    python test_hf_integration.py create_configs --model_name "$model_name" --out_dir slurm_results --flavor debugmodel
    python test_hf_integration.py submit_jobs --inp_dir slurm_results/${model_name}/debugmodel/seed_checkpoint --qos high
    while [ ! -f slurm_results/${model_name}/debugmodel/seed_checkpoint/status.txt ] || [ "$(cat slurm_results/${model_name}/debugmodel/seed_checkpoint/status.txt)" != "completed" ]; do
        echo "Waiting for seed checkpoint from ${model_name} to complete ..."
        sleep 1
    done
    python test_hf_integration.py submit_jobs --inp_dir slurm_results/${model_name}/debugmodel --qos high
    echo "================"
done

Further tasks

• Moe (handle in PR Add transformer backend (MoE) clean  #3)
  • Missing build_optimizers_with_moe_load_balancing support for MoE
  • Missing TP/PP/EP supports for MoE
• When using HF modeling, the test FSDP=2 vs FSDP=2 + PP=2, the loss and grad_norm not bitwise matching (but converging) while it is the case with Torchtitan modeling. (issue is tracked in Fix pp convergence to be bitwise #4)
• Add convergence tests to CI by doing tiny model + gloo backend (once PP is bitwise matching)
• the HF modeling has lower MFU than Torchtitan MFU
• NOTE: import torch._dynamo.config; torch._dynamo.config.cache_size_limit = 128 to avoid recomputation for graph when using torch.compile and activation checkpointing

[3outeille]

@wwwjn addresses all the issues mentioned. There is one last point I want to address (cf here) which I think provide a better user experience