[flash-linear-attention] Significant Compilation Time Discrepancy: Triton XPU (A770) vs. NVIDIA (RTX 4090)

[zhiyuan1i]

Describe the issue

Currently, I am doing continuous integration for fla.

I'm experiencing substantially slower compilation times when using Triton with Intel Arc A770 compared to NVIDIA RTX 4090 (10x+ difference in some kernel's end-to-end compilation). This significantly impacts development workflow and CI testing efficiency.

Reproduction Steps:

1. Environment: Intel Arc A770 (latest drivers) vs. NVIDIA RTX 4090
2. Clone repository: git clone https://github.com/fla-org/flash-linear-attention/
3. intall intel pytorch and fla
4. Run compilation benchmark:

export COMPILER_MODE=1 
export SKIP_TEST_CHUNK_VARLEN=1
pytest tests/ops/

Observed Behavior:

• A770 compilation takes >5-10x longer than 4090 for equivalent operations
• Noticeable delay in development iteration cycles
• Substantial impact on CI/CD pipeline efficiency

Expected Behavior:

• Comparable compilation performance between Intel and NVIDIA hardware
• Reasonable compilation times for unit testing scenarios

Impact:

1. Developer Experience: Slow compilation disrupts the edit-compile-test workflow
2. Testing Efficiency: Unit tests become bottlenecked by compilation rather than actual execution
3. Adoption Barrier: Such performance gaps may discourage potential XPU adopters

Request:
Could the Triton team:

1. Investigate this compilation performance discrepancy?
2. Share any known optimizations or workarounds for XPU compilation?
3. Consider compilation performance as a priority metric for XPU backend?

Hey, could the Triton team:

1. Check out this compilation performance discrepancy? Maybe the tiling introduces the compile speed decay
2. Share any known optimizations or workarounds for XPU compilation?
3. Think about making compilation performance a priority metric for XPU backend?

Environment details

pip list |grep triton
pytorch-triton-xpu 3.3.0
Intel Arc A770

[zhiyuan1i]

I also encountered some compilation errors and segment fault issues. I'll organize them later and create a new issue.

[alexbaden]

A770 lacks hardware features for MMA that newer hardware has (PVC, LNL, BMG). I suspect the compilation time is coming from inefficient (large) kernels. This is a known issue and, unfortunately, I'm not sure this is something we intend to improve. Do you have access to a BMG or even a PVC machine? Intel developer cloud could be a good place to start.

[zhiyuan1i]

A770 lacks hardware features for MMA that newer hardware has (PVC, LNL, BMG). I suspect the compilation time is coming from inefficient (large) kernels. This is a known issue and, unfortunately, I'm not sure this is something we intend to improve. Do you have access to a BMG or even a PVC machine? Intel developer cloud could be a good place to start.

Thank you, I will try thay if possible:)