{
  "_comment": "FMHA-specific architecture specs. Edit this file to add new GPU/dtype/pipeline support for FMHA.",
  "_note": "Common GPU hardware data (element_sizes, warp_size, warp_configs, lds_capacity_kb) lives in ../arch_specs.json. This file holds FMHA-specific capabilities, tile tables, and validation rules.",

  "architectures": {
    "gfx90a": {
      "family": "cdna2",
      "arch_tag": "ck_tile::gfx9_t",
      "supported_dtypes": ["fp16", "bf16", "fp32"],
      "supported_pipelines": ["qr", "qr_async", "qs", "qr_pagedkv", "qr_nwarp_sshuffle", "appendkv", "bwd"],
      "supports_trload": false,
      "supports_v3": false
    },
    "gfx942": {
      "family": "cdna3",
      "arch_tag": "ck_tile::gfx9_t",
      "supported_dtypes": ["fp16", "bf16", "fp32", "fp8", "fp8fp16", "fp8bf16", "fp8fp32", "bf8"],
      "supported_pipelines": ["qr", "qr_async", "qs", "qr_pagedkv", "qr_nwarp_sshuffle", "appendkv", "bwd"],
      "supports_trload": false,
      "supports_v3": false
    },
    "gfx950": {
      "family": "cdna4",
      "arch_tag": "ck_tile::gfx9_t",
      "supported_dtypes": ["fp16", "bf16", "fp32", "fp8", "fp8fp16", "fp8bf16", "fp8fp32", "bf8"],
      "supported_pipelines": ["qr", "qr_async", "qs", "qr_async_trload", "qr_async_trload_v3", "v3", "qr_pagedkv", "qr_nwarp_sshuffle", "appendkv", "bwd"],
      "supports_trload": true,
      "supports_v3": true
    },
    "gfx1100": {
      "family": "rdna3",
      "arch_tag": "ck_tile::gfx11_t",
      "supported_dtypes": ["fp16", "bf16"],
      "supported_pipelines": ["qr", "qr_pagedkv", "qr_nwarp_sshuffle", "appendkv", "bwd"],
      "supports_trload": false,
      "supports_v3": false
    },
    "gfx1201": {
      "family": "rdna4",
      "arch_tag": "ck_tile::gfx12_t",
      "supported_dtypes": ["fp16", "bf16", "fp8", "fp8bf16"],
      "supported_pipelines": ["qr", "qr_pagedkv", "qr_nwarp_sshuffle", "appendkv", "bwd"],
      "supports_trload": false,
      "supports_v3": false
    }
  },

  "supported_hdims": {
    "_comment": "hdim_q must satisfy ceil_to_qualified_tile_length() in tile_fmha_shape.hpp. Each entry is [hdim_q, hdim_v].",
    "fp16":    [[32,32], [64,64], [80,96], [96,128], [128,128], [160,160], [192,128], [192,192], [256,256]],
    "bf16":    [[32,32], [64,64], [80,96], [96,128], [128,128], [160,160], [192,128], [192,192], [256,256]],
    "fp32":    [[32,32], [48,48], [64,64], [96,128], [128,128], [192,192], [256,256]],
    "fp8":     [[64,64], [128,128], [192,128], [256,256]],
    "fp8bf16": [[64,64], [128,128], [192,128], [256,256]],
    "fp8fp32": [[128,128]],
    "bf8":     [[64,64], [128,128], [192,128], [256,256]],
    "mxfp8":   [[128,128], [256,256]],
    "mxfp4":   [[128,128], [256,256]]
  },

  "fmha_warp_tiles": {
    "_comment": "FMHA warp tile sizes [wm0, wn0, wk0] per FMHA dtype. Subset of MFMA/WMMA instructions relevant to attention.",
    "fp16":    [[32,32,16], [16,16,32]],
    "bf16":    [[32,32,16], [16,16,32]],
    "fp32":    [[16,16,16]],
    "fp8":     [[32,32,32]],
    "fp8bf16": [[32,32,32]],
    "fp8fp32": [[32,32,32]],
    "bf8":     [[32,32,32]],
    "mxfp8":   [[32,32,64]],
    "mxfp4":   [[16,16,128]]
  },

  "fmha_element_sizes": {
    "_comment": "FMHA-specific element sizes for composite dtypes not in parent arch_specs.json. Common dtypes (fp16, bf16, fp32, fp8, bf8) use ../arch_specs.json element_sizes.",
    "fp8bf16": 1,
    "fp8fp32": 1,
    "mxfp8":   1,
    "mxfp4":   1
  },

  "tile_sweep_ranges": {
    "_comment": "Block tile dimensions to sweep. Must be multiples of warp tile sizes.",
    "valid_bm0": [16, 32, 64, 128, 192, 256],
    "valid_bn0": [16, 32, 64, 96, 128, 192, 256, 384],
    "valid_bk0": [16, 32, 64, 96, 128, 256]
  },

  "k0max_map": {
    "_comment": "Maps hdim_q -> padded K-tile length. Source: tile_fmha_shape.hpp ceil_to_qualified_tile_length().",
    "32": 32, "48": 48, "64": 64, "80": 96, "96": 128,
    "128": 128, "160": 256, "192": 192, "256": 256
  },

  "lds_limits": {
    "_comment": "LDS budget in bytes per non-async FMHA pipeline. Async pipelines compute LDS dynamically.",
    "qr": 65536,
    "qs": 65536
  },

  "global_rules": {
    "hdim_192_128_no_bias_dropout": true,
    "logits_requires_no_bias": true,
    "group_mode_requires_padding": true,
    "hdim_divisible_by": 8
  },

  "defaults": {
    "tile":  [128, 64, 32, 128, 32, 128],
    "wave":  [2, 2, 1, 2, 2, 1, 1, 1, 1],
    "warp":  [32, 32, 16, 32, 32, 16, 16, 16, 16],
    "padding": [true, true, true, true],
    "block_per_cu": 1,
    "num_wave_groups": 1,
    "selection_rank": 0
  },

  "splitkv_combine": {
    "combine_bn1": 32,
    "hdims_fp16": [32, 64, 96, 128, 256],
    "hdims_fp8":  [64, 128, 256]
  },

  "batch_prefill": {
    "supported_page_sizes": [1, 16, 1024],
    "supported_kv_memory_layouts": ["vectorized", "linear"],
    "supported_kv_lookup_tables": ["vllm", "sglang"]
  },

  "bwd_tiles": {
    "_comment": "BWD dq_dk_dv tile tables. Format: [bm0, bn0, bk0, bn1, bk1, bk0max, tile6, tile7, tile8].",
    "dq_dk_dv_fp16": {
      "32_32":   [32, 128, 32,  32,  32,  32, 64,  32,  32],
      "64_64":   [32, 128, 64,  32,  64,  32, 32,  64,  64],
      "96_128":  [32, 128, 96,  32,  96,  32, 32,  96,  96],
      "128_128": [16, 128, 128, 16, 128,  16, 32, 128, 128],
      "256_256": [16,  64, 256, 16, 256,  16, 32, 256, 256]
    },
    "dq_dk_dv_extra": {
      "64_64": [
        {"tile": [32, 128, 64, 32, 64, 32, 32, 64, 64], "tag": "trload", "batch_only": false},
        {"tile": [32, 16, 64, 32, 64, 32, 16, 64, 64], "tag": "small", "batch_only": true}
      ],
      "128_128": [
        {"tile": [16, 16, 128, 16, 128, 16, 16, 128, 128], "tag": "small", "batch_only": true},
        {"tile": [16, 192, 128, 16, 128, 16, 32, 128, 128], "tag": "bn192", "batch_only": false},
        {"tile": [32, 128, 128, 32, 128, 32, 32, 128, 128], "tag": "trload", "batch_only": false}
      ]
    },
    "dot_do_o_hdims": [32, 64, 96, 128, 256],
    "convert_dq_hdims": [32, 64, 96, 128, 256],
    "convert_dq_tile_groups": {"32": 1, "64": 1, "96": 1, "128": 1, "256": 1},
    "pad_combos": [["f","f"], ["f","t"], ["f","8"], ["t","f"], ["t","t"], ["t","8"], ["8","8"]],
    "extra_pad_combos": [["f","f"], ["8","8"]],
    "dropouts": ["no", "dropout_wg16", "dropout_wg16_storerandval"],
    "small_dropouts": ["no"]
  },

  "bwd_wave_warp": {
    "_comment": "BWD wave/warp lookup. Key: 'bm0_bn0_bk0_trload'. Value: {wave: 9-tuple, warp_k1: int}.",
    "16_16_128_t":   {"wave": [1,1,1,1,1,1,1,1,1], "warp_k1": 16},
    "16_64_256_f":   {"wave": [1,4,1,4,1,1,1,4,1], "warp_k1": 16},
    "16_128_128_f":  {"wave": [1,4,1,4,1,1,1,4,1], "warp_k1": 16},
    "16_192_128_t":  {"wave": [1,4,1,4,1,1,1,4,1], "warp_k1": 16},
    "32_16_64_t":    {"wave": [1,1,1,1,1,1,1,1,1], "warp_k1": 16},
    "32_128_32_f":   {"wave": [1,4,1,4,1,1,2,2,1], "warp_k1": 16},
    "32_128_64_f":   {"wave": [1,4,1,4,1,1,1,4,1], "warp_k1": 16},
    "32_128_64_t":   {"wave": [1,4,1,4,1,1,1,4,1], "warp_k1": 32},
    "32_128_96_f":   {"wave": [1,4,1,4,1,1,2,2,1], "warp_k1": 16},
    "32_128_128_t":  {"wave": [1,4,1,4,1,1,1,4,1], "warp_k1": 32},
    "64_128_32_f":   {"wave": [2,4,1,4,1,1,2,4,1], "warp_k1": 16},
    "64_128_64_f":   {"wave": [2,4,1,4,1,1,2,4,1], "warp_k1": 16},
    "64_128_128_f":  {"wave": [2,4,1,4,1,1,2,4,1], "warp_k1": 16}
  }
}