project_6/muh/schema/three_way_partition.yaml

# muh schema for three_way_partition
# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_three_way_partition.cuh
# Generated by muh/extract.py

algorithm: three_way_partition
source: cub/cub/device/dispatch/tuning/tuning_three_way_partition.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
  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
  input_size:
    type: int
    range:
      - 1
      - 1024
    step: 1
    note: unknown_range
  offset_size:
    type: int
    range:
      - 1
      - 1024
    step: 1
    note: unknown_range
reference_tunings:
  sm80:
    -
      threads: 256
      items: 12
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      threads: 256
      items: 11
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      threads: 224
      items: 11
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      threads: 128
      items: 10
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
  sm90:
    -
      threads: 256
      items: 12
      load_algorithm: BLOCK_LOAD_DIRECT
    -
      threads: 256
      items: 12
      load_algorithm: BLOCK_LOAD_DIRECT
    -
      threads: 320
      items: 12
      load_algorithm: BLOCK_LOAD_DIRECT
    -
      threads: 384
      items: 7
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      threads: 128
      items: 7
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      threads: 256
      items: 24
      load_algorithm: BLOCK_LOAD_DIRECT
    -
      threads: 640
      items: 24
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      threads: 256
      items: 23
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      threads: 256
      items: 18
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      threads: 256
      items: 11
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      threads: 128
      items: 7
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      threads: 128
      items: 7
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
  sm100:
    -
      items: 12
      threads: 256
      load_algorithm: BLOCK_LOAD_DIRECT
    -
      items: 14
      threads: 288
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      items: 11
      threads: 512
      load_algorithm: BLOCK_LOAD_DIRECT
    -
      items: 10
      threads: 256
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      items: 20
      threads: 768
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      items: 20
      threads: 768
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      items: 15
      threads: 768
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
    -
      items: 14
      threads: 320
      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
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
[MUH] Bootstrap muh toolchain — extract/parse/gen_yaml/gen_patch + baseline.muh 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 2026-07-30 10:39:06 +00:00			`# muh schema for three_way_partition`
			`# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_three_way_partition.cuh`
			`# Generated by muh/extract.py`

			`algorithm: three_way_partition`
			`source: cub/cub/device/dispatch/tuning/tuning_three_way_partition.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`
			`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`
			`input_size:`
			`type: int`
			`range:`
			`- 1`
			`- 1024`
			`step: 1`
			`note: unknown_range`
			`offset_size:`
			`type: int`
			`range:`
			`- 1`
			`- 1024`
			`step: 1`
			`note: unknown_range`
			`reference_tunings:`
			`sm80:`
			`-`
			`threads: 256`
			`items: 12`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`threads: 256`
			`items: 11`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`threads: 224`
			`items: 11`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`threads: 128`
			`items: 10`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`sm90:`
			`-`
			`threads: 256`
			`items: 12`
			`load_algorithm: BLOCK_LOAD_DIRECT`
			`-`
			`threads: 256`
			`items: 12`
			`load_algorithm: BLOCK_LOAD_DIRECT`
			`-`
			`threads: 320`
			`items: 12`
			`load_algorithm: BLOCK_LOAD_DIRECT`
			`-`
			`threads: 384`
			`items: 7`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`threads: 128`
			`items: 7`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`threads: 256`
			`items: 24`
			`load_algorithm: BLOCK_LOAD_DIRECT`
			`-`
			`threads: 640`
			`items: 24`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`threads: 256`
			`items: 23`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`threads: 256`
			`items: 18`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`threads: 256`
			`items: 11`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`threads: 128`
			`items: 7`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`threads: 128`
			`items: 7`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`sm100:`
			`-`
			`items: 12`
			`threads: 256`
			`load_algorithm: BLOCK_LOAD_DIRECT`
			`-`
			`items: 14`
			`threads: 288`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`items: 11`
			`threads: 512`
			`load_algorithm: BLOCK_LOAD_DIRECT`
			`-`
			`items: 10`
			`threads: 256`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`items: 20`
			`threads: 768`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`items: 20`
			`threads: 768`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`items: 15`
			`threads: 768`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`-`
			`items: 14`
			`threads: 320`
			`load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE`
			`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`