Add LLM-driven tile search for CUTLASS SYCL kernels#22
Conversation
Code reviewNo issues found. Checked for bugs and CLAUDE.md compliance. 🤖 Generated with Claude Code - If this code review was useful, please react with 👍. Otherwise, react with 👎. |
danielfleischer
left a comment
danielfleischer
left a comment
There was a problem hiding this comment.
LGTM, made some comments.
There was a problem hiding this comment.
Why do we do these algebraic validations? it's not very AI
There was a problem hiding this comment.
because I saw number of time GPT proposing some weird NVIDIA style configuration, which will fail on our XPU. So instead of waiting time, I’m discarding those.
| logger = logging.getLogger(__name__) | ||
|
|
||
|
|
||
| def _setup_dspy(config: Config) -> None: |
There was a problem hiding this comment.
Why not put it inside TileTuningAgent or that module?
There was a problem hiding this comment.
_setup_dspy is a global configuration (env vars + dspy.configure()), not agent-specific logic. So I would keep it here. I wanted to have it one place.
2 participants
Add tile configuration tuning for CUTLASS SYCL kernels (GEMM, Grouped GEMM, MoE GEMM, Flash Attention V2) using a propose-validate-benchmark loop driven by DSPy/LLM
added validators enforce Intel Xe DPAS constraints (atom shapes, subgroup limits, SLM capacity)