[NEW] Llama3.2 weight converters 🦙

[TJ-Solergibert]

Hi!

In this branch I reintroduce & update to the current main branch the Llama model & conversion scripts to support Llama3.1 and Llama3.2 1B&3B models.
The main changes are the following:

1. Deleted Flash Attention RoPEs as they don't support rope scaling of Llama3
2. Copied + Cleaning transformers LlamaRotaryEmbedding layer. Now this will be the only class in llama.py. I think it shouldn't break generations for the inference case WITHOUT LlamaConfig.rope_interleaved = True in CausalSelfAttention.forward, are there any tests?
3. Added @eliebak fixes (resuming checkpoint without lr schedule or optimizer state #253) to just load weights from a checkpoint. I would suggest adding a config.optimizer.finetuning flag in order to (True) just load the weights or (False) Load weights, optimizer & LR Scheduler instead of config.checkpoints.load_optimizer & config.checkpoints.load_lr_scheduler
4. Switched from flash_attn_varlen_func to flash_attn_func as the later achieves greater performance. Keep in mind that we aren't using any feature of the varlen funct so it's recommended to stick with flash_attn_func
5. Should we depreciate LlamaConfig.rope_interleaved ? It was useful for training when using FlashAttention RoPEs and now seems to be used also in the inference code. IMO we should unify all 3 cases (Training, inference with rope_interleaved & inference without rope interleaved) within a single RoPE

Results

You can run the conversions & generations tests using the scripts in tools/converters. As I already mentioned in the previous PR (#174), despite we need at least 1 GPU (To init the ParallelContext) we are running the conversion with the CPU.

1. Generate HF predictions with the HF Hub model

torchrun --nproc-per-node 1 tools/converters/delete/generate_hf_predictions.py --pretrained-model-name-or-path meta-llama/Llama-3.2-3B

2. Convert the HF model to Nanotron

torchrun --nproc-per-node 1 tools/converters/convert_hf_to_nanotron.py --nanotron-checkpoint-path checkpoints/nanotron_pretrained_checkpoints/Nanotron-Llama-3.2-3B --pretrained-model-name-or-path meta-llama/Llama-3.2-3B

3. Generate predictions with the nanotron converted model

torchrun --nproc-per-node 1 tools/converters/delete/generate_nanotron_predictions.py --tp 1 --nanotron-checkpoint-path checkpoints/nanotron_pretrained_checkpoints/Nanotron-Llama-3.2-3B

4. Convert the model back to HF

torchrun --nproc-per-node 1 tools/converters/convert_nanotron_to_hf.py --nanotron-checkpoint-path checkpoints/nanotron_pretrained_checkpoints/Nanotron-Llama-3.2-3B --hugging-face-checkpoint-path checkpoints/huggingface_converted/Converted-Nanotron-Llama-3.2-3B

5. Generate predictions using the converted back HF model

torchrun --nproc-per-node 1 tools/converters/delete/generate_hf_predictions.py --pretrained-model-name-or-path checkpoints/huggingface_converted/Converted-Nanotron-Llama-3.2-3B

As can be seen from the following table, we observe slightly differences between the 2 backends. Those differences are produced by the QKV projections in the CausalSelfAttention layer (Nanotron computes them in a single GEMM vs 3 different GEMMs in HF) and the LayerNorm layer is different (Nanotron is using a optimized one from FlashAttention vs Basic PyTorch LayerNorm in HF). Also note that the differences increase if we use TP which is totally expected as the sizes of the GEMMs are different, triggering different GEMM algorithms.

Experiment	Backend	Size	TP	Accuracy
OG HF	HF	3	1	0.73046875
OG HF --> Nanotron	Nanotron	3	1	0.7265625
OG HF --> Nanotron --> HF	HF	3	1	0.73046875
OG HF --> Nanotron	Nanotron	3	2	0.703125
OG HF --> Nanotron	Nanotron	3	4	0.65234375

To run the Nanotron generations with different TP sizes:

torchrun --nproc-per-node 2 tools/converters/delete/generate_nanotron_predictions.py --tp 2 --nanotron-checkpoint-path checkpoints/nanotron_pretrained_checkpoints/Nanotron-Llama-3.2-3B
torchrun --nproc-per-node 4 tools/converters/delete/generate_nanotron_predictions.py --tp 4 --nanotron-checkpoint-path checkpoints/nanotron_pretrained_checkpoints/Nanotron-Llama-3.2-3B

TODO (Preferably in other PRs):

• Add docs
• Edit examples/config_llama3.2-3B.yaml data_stages.data.dataset.dataset_folder
• Remove nanotron/tools/converters/delete/generate_hf_predictions.py & nanotron/tools/converters/delete/generate_nanotron_predictions.py scripts
• Integrate Liger kernels for apply_rotary_pos_emb
• Switch to SDPA instead of FA2
• Reduce number of transposes in CausalSelfAttention.forward