9b21a13119
Pipeline: 1. extract.py: Parses all 26 CCCL tuning_*.cuh → 26 YAML schemas in muh/schema/ 2. parse.py: .muh file parser with extends-inheritance + schema validation 3. gen_yaml.py: .muh → computility-run.yaml (verified: matches competition reference) 4. gen_patch.py: .muh → vllm kernel unified diff patches (6 algorithm mappings) 5. baseline.muh: Competition reference config, all tuning values pending BI-V100 benchmarks Schemas extracted: 26 algorithms, 8-19 params each, SM75/80/90/100 reference tunings Priority mapping: reduce→attention, topk→sampling, scan→paged_attention, transform→activations, batch_memcpy→KV_cache, for→RoPE Tested: extract→parse→validate→gen_yaml→gen_patch full pipeline passes
256 lines
5.3 KiB
YAML
256 lines
5.3 KiB
YAML
# muh schema for scan
|
|
# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_scan.cuh
|
|
# Generated by muh/extract.py
|
|
|
|
algorithm: scan
|
|
source: cub/cub/device/dispatch/tuning/tuning_scan.cuh
|
|
parameters:
|
|
threads_per_block:
|
|
type: int
|
|
range:
|
|
- 32
|
|
- 1024
|
|
step: 32
|
|
items_per_thread:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 32
|
|
step: 1
|
|
load_algorithm:
|
|
type: enum
|
|
values:
|
|
- BLOCK_LOAD_DIRECT
|
|
- BLOCK_LOAD_VECTORIZE
|
|
- BLOCK_LOAD_TRANSPOSE
|
|
- BLOCK_LOAD_WARP_TRANSPOSE
|
|
- BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
|
|
- BLOCK_LOAD_STRIPED
|
|
load_modifier:
|
|
type: enum
|
|
values:
|
|
- LOAD_DEFAULT
|
|
- LOAD_CA
|
|
- LOAD_CG
|
|
- LOAD_CS
|
|
- LOAD_CV
|
|
- LOAD_LDG
|
|
store_algorithm:
|
|
type: enum
|
|
values:
|
|
- BLOCK_STORE_DIRECT
|
|
- BLOCK_STORE_WARP_TRANSPOSE
|
|
- BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED
|
|
- BLOCK_STORE_STRIPED
|
|
scan_algorithm:
|
|
type: enum
|
|
values:
|
|
- BLOCK_SCAN_RAKING
|
|
- BLOCK_SCAN_RAKING_MEMOIZE
|
|
- BLOCK_SCAN_WARP_SCANS
|
|
lookback_delay.kind:
|
|
type: enum
|
|
values:
|
|
- no_delay
|
|
- fixed_delay
|
|
- exponential_backoff
|
|
- exponential_backoff_jitter
|
|
- exponential_backoff_jitter_window
|
|
- exponential_backon_jitter_window
|
|
- exponential_backon_jitter
|
|
- exponential_backon
|
|
lookback_delay.delay:
|
|
type: int
|
|
range:
|
|
- 0
|
|
- 2000
|
|
step: 50
|
|
lookback_delay.l2_write_latency:
|
|
type: int
|
|
range:
|
|
- 0
|
|
- 2000
|
|
step: 50
|
|
reduce_and_scan_warps:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 8
|
|
step: 1
|
|
lookahead_items_per_thread:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 16
|
|
step: 1
|
|
lookahead_stages:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 1024
|
|
step: 1
|
|
note: unknown_range
|
|
block_idx_stages:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 1024
|
|
step: 1
|
|
note: unknown_range
|
|
delay_constructor:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 1024
|
|
step: 1
|
|
note: unknown_range
|
|
input_value_size:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 1024
|
|
step: 1
|
|
note: unknown_range
|
|
input_value_alignment:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 1024
|
|
step: 1
|
|
note: unknown_range
|
|
output_value_size:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 1024
|
|
step: 1
|
|
note: unknown_range
|
|
output_value_alignment:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 1024
|
|
step: 1
|
|
note: unknown_range
|
|
accum_size:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 1024
|
|
step: 1
|
|
note: unknown_range
|
|
accum_alignment:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 1024
|
|
step: 1
|
|
note: unknown_range
|
|
offset_size:
|
|
type: int
|
|
range:
|
|
- 1
|
|
- 1024
|
|
step: 1
|
|
note: unknown_range
|
|
reference_tunings:
|
|
sm75:
|
|
-
|
|
threads: 128
|
|
items: 7
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
load_modifier: LOAD_DEFAULT
|
|
sm80:
|
|
-
|
|
threads: 320
|
|
items: 14
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
-
|
|
threads: 352
|
|
items: 16
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
-
|
|
threads: 320
|
|
items: 12
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
-
|
|
threads: 288
|
|
items: 22
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
-
|
|
threads: 288
|
|
items: 8
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
-
|
|
threads: 384
|
|
items: 12
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
-
|
|
threads: 640
|
|
items: 24
|
|
load_algorithm: BLOCK_LOAD_DIRECT
|
|
store_algorithm: BLOCK_STORE_DIRECT
|
|
sm90:
|
|
|
|
sm100:
|
|
-
|
|
items: 18
|
|
threads: 512
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
load_modifier: LOAD_DEFAULT
|
|
-
|
|
items: 14
|
|
threads: 384
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
load_modifier: LOAD_CA
|
|
-
|
|
items: 13
|
|
threads: 512
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
load_modifier: LOAD_DEFAULT
|
|
-
|
|
items: 13
|
|
threads: 288
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
load_modifier: LOAD_CA
|
|
-
|
|
items: 22
|
|
threads: 384
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
load_modifier: LOAD_DEFAULT
|
|
-
|
|
items: 19
|
|
threads: 416
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
load_modifier: LOAD_CA
|
|
-
|
|
items: 23
|
|
threads: 416
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
load_modifier: LOAD_DEFAULT
|
|
-
|
|
items: 22
|
|
threads: 320
|
|
load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
|
|
store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
|
|
load_modifier: LOAD_DEFAULT
|
|
bi_v100:
|
|
status: pending_benchmark
|
|
note: Run muh benchmark on Iluvatar BI-V100 to fill these values
|
|
threads_per_block: TBD
|
|
items_per_thread: TBD
|