[Llamma4] Chunked Attention

[vasqu]

System Info

• transformers version: 4.52.0.dev0 (around commit 8bbcdf5)
• Platform: Linux-6.8.0-111057-tuxedo-x86_64-with-glibc2.35
• Python version: 3.10.12
• Huggingface_hub version: 0.30.1
• Safetensors version: 0.4.3
• Accelerate version: 0.34.2
• Accelerate config: not found
• DeepSpeed version: not installed
• PyTorch version (GPU?): 2.6.0+cu124 (True)
• Tensorflow version (GPU?): 2.15.1 (True)
• Flax version (CPU?/GPU?/TPU?): 0.7.0 (cpu)
• Jax version: 0.4.13
• JaxLib version: 0.4.13
• Using distributed or parallel set-up in script?:
• Using GPU in script?:
• GPU type: NVIDIA GeForce RTX 3080 Ti Laptop GPU

Who can help?

@ArthurZucker @winglian (fyi)

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Because I'm GPU poor, I modified llama4 to only have one layer and a lower hidden size.

Rough script:

import torch

from transformers import AutoConfig, AutoTokenizer
from transformers.models.llama4.modeling_llama4 import Llama4ForCausalLM


config = AutoConfig.from_pretrained("meta-llama/Llama-4-Scout-17B-16E-Instruct").get_text_config()

# modify config for debugging
config._attn_implementation = "eager"  # or "flex_attention", possibly also "sdpa"
config.hidden_size = 128
config.num_hidden_layers = 1
config.attention_chunk_size = 3 # causes eager issues, leaving default causes issues in flex attention

# some dummy data
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-4-Scout-17B-16E-Instruct")
tokenizer.padding_side = "left"  # irrelevant tbh
input_text = ["What are we having for dinner?", "How are you?"]
input_ids = tokenizer(input_text, padding=True, return_tensors="pt").to("cuda")

# init module, half precision to save on vram 
test_module = Llama4ForCausalLM(config).to("cuda", torch.bfloat16)

# simple forward pass
test_module.forward(**input_ids)

This can cause various issues, e.g.

• In eager: RuntimeError: The size of tensor a (8) must match the size of tensor b (2) at non-singleton dimension 0
• In flex: ValueError: block_mask was created for block_mask.shape=(2, 1, 8, tensor(8192, device='cuda:0')) but got q_len=8 and kv_len=8. (...) - looks like more fixes for post-training llama4 #37329 (comment)

Expected behavior

Chunked attention doesn't seem to be correctly handled atm. A lots of code does not enter this territory because of the fairly long context to even go over the required chunk size.

[vasqu]

Also as a note:

• It has to be torch==2.6.0, previous versions (2.5.1) can handle this somewhat for flex attention (when disabling mask creation compilation - another bug introduced).
• The purpose/usage behind chunked attention is unclear to me. It might also not handle padding correctly... (but that's only on first glance).

[ArthurZucker]

Chunked attention enable 10M context length! It discards unused cache on the go!

[vasqu]

The attention mask reminded me of packed sequences 😄 But that sounds nice!

I forgot to mention that I also modified the mask padding (pls see vasqu@5f9b658). (Otherwise smaller sequences will crash on flex attention)

[overwindows]

Same issue here

when running the sample code from the LLaMA 4 release blog, and even without using Flex Attention (e.g., using eager or sdpa), it still throws an error if the input length exceeds 8K.

chunked_attention_mask = chunked_attention_mask & attention_mask RuntimeError: The size of tensor a (8192) must match the size of tensor b (16395) at non-singleton dimension 1

[ArthurZucker]

this should be fixed with latest patch release!