FMS Acceleration for Fused Operations and Kernels

This library contains fused operations and custom kernels, to be expanded over time. Currently it contains the following:

1. Fused operations and kernels extracted from unsloth.
   • Low-Rank Adapter Fused Operations
   • Fast RoPE Triton Kernels
   • Fast RMS LayerNorm Triton Kernels
   • Fast Cross Entropy Triton Kernels

Plugins

Supported DataType Settings

Compatibility Matrix with Mixed Precision

NOTE: this chart is also a good reference for supported types, even for the non-FOAK case.

Code Extracted from Unsloth

Notes on the extraction of code from unsloth:

• While unsloth is released under Apache 2.0, there are comments indicating some exceptions strewn throughout the code base, see an example here.

  it would require a commercial license if used to run on more than 4 GPUs ...
  

• These exceptions appear to be located around the trainer improvements, see another example here.
• These exceptions appear around Feb 2024 Release; any code that appears in any file where such exceptions occur is not extracted.
• Instead in its place, we have adopted a different approach; we adopt the approach of model patching, as opposed unsloths' approach to rewrite the model. Our approach is novel and completely rewritten from scratch.
• We have also enabled dropout on the lora fused operations.
• All extracted code appears before the Feb 2024 Release.
• In the table below we record what was extracted, and the exact commit from which it was taken.

Supported Models

Adding Support For A New Model

It is realtively easy by following an existing template, in what follows we use GraniteForCausalLM as an example.

• implement a get_mp_rules for the new model, which returns a list of ModelPatcherRule.
• logic that needs to be changed is the various classes that the rules are triggered on. Import the various module classes likes so:

  from transformers.models.granite.modeling_granite import ( 
      GraniteAttention,
      GraniteMLP,
      GraniteRMSNorm,
  )
  

• replace the classes appropriately in various locations in ModelPatcherRule. In particular the ModelPatcherTrigger portions of it. Name rule_id appropriately.

  ModelPatcherRule(
      rule_id="granite-rms",
      trigger=ModelPatcherTrigger(check=GraniteRMSNorm),
      forward=fast_rms_layernorm,
  )
  

Running Liger Kernel Benchmarks

Using the scenarios-liger.yaml, this will run full fine tuning, lora peft, autoGPTQ lora peft, and bits-and-bytes lora peft with the triton kernels (Fast RMS, RoPE, CrossEnt) as a base and then run with the liger kernel for LigerFusedLinearCrossEntropy as well as Fast RMS, RoPE to compare results. It only runs against mistral and llama models.

The benchmarks were ran separately for each num_gpu entry; they can be run together in a single command, but this is more efficient.

tox -e run-benches -- 1 "4 8 16 32" benchmark_outputs_1 scenarios-liger.yaml none
tox -e run-benches -- 2 "8 16 32 64" benchmark_outputs_2 scenarios-liger.yaml none
tox -e run-benches -- 4 "16 32 64 128" benchmark_outputs_3 scenarios-liger.yaml none

Known Issues

• MixedPrecision --fp16 or --bf16 should be used with fast_lora.
• fast_lora has issues with FSDP V1 with the peft style of FSDP wrapping.
  • This is because the adapter's forward functions are bypassed in the fused ops.
  • For AutoGPTQ/QLoRA this is addressed by distributing the adapters using DDP so they will be unsharded in time for the fused ops.
• fast_rope_embeddings does not work with postion_ids, it seems like HF has depracated passing these ids into the rope embedding methods.