[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

muh/schema/rle_encode.yaml

@@ -0,0 +1,155 @@
+# muh schema for rle_encode
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_rle_encode.cuh
+# Generated by muh/extract.py
+
+algorithm: rle_encode
+source: cub/cub/device/dispatch/tuning/tuning_rle_encode.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
+  length_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+reference_tunings:
+  sm80:
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 13
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 13
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 224
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+  sm90:
+    -
+      threads: 256
+      items: 13
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 128
+      items: 22
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 192
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 19
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+  sm100:
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_CA
+    -
+      threads: 224
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_CA
+    -
+      threads: 224
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      nominal_4B_items_per_thread: 6
+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