File size: 2,724 Bytes
9d3ee31 | 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 77 78 79 80 81 82 | {
"container": "trained_susv_wave_gamma_chunked_v1",
"decode_entry": "decode_expert_layer_v3t / decode_expert_v3t (decode.py)",
"fields": "trellis|su|sv|wave_gamma",
"susv_mode": "continuous fp16, Wscale absorbed in sv (replaces int8 +-1 signs)",
"wave_gamma": {
"len": "Mb+Nb (transpose-invariant; 160 at 512x2048 pads)",
"index": "wave_index_map(Mb,Nb) in decode.py \u2014 LAST wavefront that wrote each tile",
"granularity": "constant per 16x16 tile; constant along each encoder wavefront"
},
"op_order": "recons -> fp16 cast -> wave_gamma (BEFORE both Hadamards) -> H_n -> *su -> T -> H_m -> *sv -> T -> crop",
"tlut": {
"file": "packed/experts/codebook.safetensors",
"structure": "exact scaled integer lattice: global_scale * levels",
"global_scale": 0.8963924050331116,
"integer_levels": [
-5,
-4,
-3,
-2,
-1,
0,
1,
2,
3,
4
],
"n_levels": 10,
"ratios_exactly_integer": true
},
"kernel_contract": {
"adds": "per-weight ops remain 100% add/subtract on the exact lattice (codes + tlut unchanged)",
"stage_1_wave_gamma": "scale each 16x16 tile's partial sums by gamma[wave(tile)] (fp16) BEFORE cross-tile accumulation; ~1 multiply per 16 weights",
"stage_2_susv": "un-rotation sign-flips become per-dim fp16 multiplies (su over n, sv over m); n+m mults/GEMV",
"gate_guidance": "codebook-only lattice gates MUST also key on this file's container field"
},
"cb_params": {
"K": 1.5,
"L": 16,
"V": 8,
"tlut_bits": 15,
"decode_mode": "quantlut_sym",
"td_x": 16,
"td_y": 16
},
"ne_tier": {
"live_set": "packed/ne/L00-L39.safetensors (zero-padded, all layers)",
"codec": "canon_rht_bitshift_trellis_intlattice",
"cb_params": {
"K": 4,
"L": 16,
"V": 2,
"tlut_bits": 9,
"decode_mode": "quantlut_sym",
"td_x": 16,
"td_y": 16
},
"keys": "<tensor>|{trellis,SU,SV,Wscale} (int8 sign SU/SV + scalar fp16 Wscale)",
"metadata": "per-file: codec + cb_params + dims \u2014 NO manifest key",
"tlut": {
"file": "packed/ne/tlut.safetensors",
"shape": [
512,
2
],
"sha256_16": "41a6b4f5ae0ca5ab"
},
"decode_entry": "decode_ne_shard (decode.py; dispatches on codec metadata)"
},
"head_tier": {
"files": "packed/head/head_c{0..7}of8.safetensors",
"codec": "int5g64_packed",
"keys": "LMHEADCHUNK:{r0}:{r1}|{qp,gscale} (+ optional |prot_rows,|prot_dense)",
"decode_entry": "decode_head (decode.py)"
},
"embed_tier": {
"file": "packed/ne/embed_int4.safetensors",
"format": "affine int4 asym g64 + exact-overwrite exceptions (encoder-GPU FMA boundary elements)",
"keys": "q_packed|mn|mx|exc_idx|exc_bits",
"decode_entry": "decode_embed(packed_dir, bits=4) (decode.py, pure numpy)"
}
} |