File size: 4,885 Bytes
0f775e2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 | """TaskSpec — everything that varies between kernel-generation tasks.
Scaling this lane to ~150 kernels means writing ONLY the parts that are genuinely per-kernel. Everything
else (Dockerfile, measure.py, the grader with its anti-cheat, test.sh, task.toml, RUN.md, the precision +
faithfulness policy, the grading-transparency section) is generated from these fields by `build.py`.
A new task is therefore: a reference implementation, an input generator, a FLOP/byte formula, shape lists,
and a few paragraphs of prose. Everything in this file is data; nothing here runs on the GPU.
"""
from dataclasses import dataclass, field
@dataclass
class TaskSpec:
# ---- identity -------------------------------------------------------------------------------
name: str # directory + task name, e.g. "kda-forward"
title: str # instruction.md H1, e.g. "Write a fast Kimi Delta Attention ... kernel"
blurb: str # one-paragraph what/why, used in task.toml + RUN.md
keywords: list[str] = field(default_factory=list)
# ---- the graded entry point ------------------------------------------------------------------
module: str = "" # file the agent edits, e.g. "kda.py"
func: str = "" # function the grader imports, e.g. "kda_forward"
signature: str = "" # e.g. "kda_forward(q, k, v, g, beta, scale=None)"
returns_doc: str = "" # markdown describing the return contract
# ---- code the generator embeds verbatim ------------------------------------------------------
reference_src: str = "" # the reference implementation (also embedded into the grader)
reference_imports: str = "" # imports the reference needs, e.g. "import torch\nfrom einops import ..."
make_inputs_src: str = "" # def _mk(*shape, seed) -> tuple of tensors
flops_src: str = "" # def canonical_work(*shape) -> int (or bytes, for the GB/s metric)
flops_formula: str = "" # OPTIONAL: the formula PRINTED in instruction.md. If left empty it is
# scraped from flops_src's last `return`, which requires canonical_work
# to come last and end in a single-expression return. Set this
# explicitly whenever the work count uses helpers or several statements,
# otherwise the instruction can show a partial/meaningless formula.
# ---- grading ---------------------------------------------------------------------------------
metric: str = "TFLOP/s" # "TFLOP/s" | "GB/s" | "tokens/s"
compare: str = "tensor" # "tensor" | "tuple" | "rowwise"
tol: float = 2e-2
tuple_names: tuple = () # for compare="tuple", e.g. ("dq","dk","dv","dg","dbeta")
row_pass: float = 0.98 # for compare="rowwise"
grader_shapes: list = field(default_factory=list)
measure_shapes: list = field(default_factory=list)
measure_quick_shapes: list = field(default_factory=list)
correct_shapes: list = field(default_factory=list)
shape_names: tuple = () # for pretty-printing, e.g. ("B","T","H","K","V")
# ---- prose (markdown, dropped into instruction.md) --------------------------------------------
spec_md: str = "" # "## The computation (this is the exact spec)" body
contract_md: str = "" # the fixed-contract table + notes
regime_md: str = "" # the graded shape regime paragraph
perf_md: str = "" # "## Where the performance comes from" body
precision_md: str = "" # task-specific opening of the precision section
correctness_md: str = "" # task-specific wording of the correctness gate
# ---- environment ------------------------------------------------------------------------------
base_image: str = "pytorch/pytorch:2.11.0-cuda12.8-cudnn9-devel" # torch 2.11 + triton 3.6
pip_extra: str = "einops nvidia-cutlass-dsl"
gpus: int = 1
agent_timeout_sec: float = 14400.0
verifier_timeout_sec: float = 1800.0
memory_mb: int = 65536
def validate(self):
assert self.name and self.module and self.func, "name/module/func are required"
assert self.compare in ("tensor", "tuple", "rowwise"), self.compare
assert self.metric in ("TFLOP/s", "GB/s", "tokens/s"), self.metric
assert self.grader_shapes and self.correct_shapes, "shape lists are required"
if self.compare == "tuple":
assert self.tuple_names, "compare='tuple' needs tuple_names"
for s in self.grader_shapes + self.correct_shapes + self.measure_shapes:
assert len(s) == len(self.shape_names), f"shape {s} vs names {self.shape_names}"
return self
|