Upload 2 files

hexstate_requantize.py

@@ -5,10 +5,9 @@ HexState GGUF Re-Quantizer — GGUF-to-GGUF Q2_K quantization.
 Reads a source GGUF (F16/BF16/F32), copies all metadata verbatim,
 and re-quantizes eligible weight tensors to Q2_K.  Fold-interleave is ON
 by default: each Q2_K block is order-stat paired (k-th smallest at k,
-k-th largest at k+128) and fold_perm.* I16 tensors are stored for
-FoldLlama, which inverts the perm at matmul (no requant).  Pass
---no-fold-interleave for a plain Q2_K GGUF.  Optional --fold-basis is a
-shared dimension perm (stock llama, does not hit Fold RMSE).
+k-th largest at k+128) then written as ordinary 84-byte Q2_K — stored
+order is decode order, no sidecar.  Pass --no-fold-interleave for
+identity-layout Q2_K.  Optional --fold-basis is a shared dimension perm.
 
 This bypasses the tokenizer parsing problem entirely — the source GGUF
 (from llama.cpp's convert_hf_to_gguf.py) has correct metadata.
@@ -20,7 +19,6 @@ Usage:
 import struct
 import sys
 import time
-import tempfile
 import os
 import io
 import ctypes
@@ -757,7 +755,7 @@ def apply_fold_interleave_importance(importance, perm_flat, block_size=QK_K):
 
 
 def _stream_quantize(fin, fout, ti, abs_offset, kind, fb_perms, imatrix_data,
-                     use_hpc, fold_interleave=False, perm_file=None):
+                     use_hpc, fold_interleave=False):
     """Quantize one tensor in row chunks. kind: 'q2k' | 'q4' | 'q8'.
     Returns (n_out_bytes, rmse_or_None). Copies raw if source is already quant.
     """
@@ -801,8 +799,6 @@ def _stream_quantize(fin, fout, ti, abs_offset, kind, fb_perms, imatrix_data,
 
         if kind == 'q2k' and fold_interleave:
             perm_u8 = apply_fold_interleave(f32)
-            if perm_file is not None:
-                perm_file.write(np.ascontiguousarray(perm_u8, dtype='<i2').tobytes())
             if imp is not None:
                 imp = apply_fold_interleave_importance(imp, perm_u8)
 
@@ -1351,7 +1347,7 @@ def main():
         print("Usage: python3 hexstate_requantize.py <input.gguf> <output.gguf>"
               " [--keep-metadata] [--imatrix FILE] [--keep-embd] [--q2all]"
               " [--fold-basis] [--no-fold-interleave]")
-        print("  Fold interleave is ON by default (Q2_K + fold_perm.* for FoldLlama).")
+        print("  Fold interleave is ON by default (in-block Q2_K layout, no extra tensors).")
         print("  HEX_CHUNK_ELEMS  max f32 elements per tensor chunk (default 2000000)")
         sys.exit(1)
 
@@ -1386,7 +1382,7 @@ def main():
     if q2all:
         print("  ║  Mode: --q2all  ALL eligible tensors → Q2_K (test mode)      ║")
     if fold_interleave:
-        print("  ║  Fold-interleave: ON  (Q2_K + perm tensors, FoldLlama)       ║")
+        print("  ║  Fold-interleave: ON  (in-block Q2_K, same file size)        ║")
     if fold_basis:
         print("  ║  Fold-basis: ON  (shared dim perm, optional)                 ║")
     print(f"  ║  Chunk: {_chunk_max_elems():<7d} f32 elems/tensor (HEX_CHUNK_ELEMS)     ║")
@@ -1581,18 +1577,6 @@ def main():
             })
             out_data_offset += out_size
             out_data_offset = align_offset(out_data_offset)
-            if fold_interleave and out_type == GGML_TYPE_Q2_K:
-                psize = int(ti['n_elements']) * 2
-                out_tensor_infos.append({
-                    'name': 'fold_perm.' + ti['name'],
-                    'n_dims': 1,
-                    'dims': [int(ti['n_elements'])],
-                    'type': GGML_TYPE_I16,
-                    'offset': out_data_offset,
-                    'data_size': psize,
-                })
-                out_data_offset += psize
-                out_data_offset = align_offset(out_data_offset)
 
         # ── Update KV pairs ──
         updated_kv = []
@@ -1687,14 +1671,6 @@ def main():
                 else:
                     updated_kv.append((key, vtype, raw_value))
 
-        if fold_interleave:
-            val = b'true'
-            updated_kv.append((
-                'hexstate.fold_interleave', 8,
-                struct.pack('<Q', len(val)) + val,
-            ))
-            print("  KV hexstate.fold_interleave = true")
-
         # ── Write output GGUF ──
         print("  Writing output GGUF...")
         with open(output_path, 'wb') as fout:
@@ -1768,13 +1744,10 @@ def main():
                     total_quant_bytes += nbytes
 
                 elif plan:
-                    perm_file = None
-                    if fold_interleave:
-                        perm_file = tempfile.SpooledTemporaryFile(max_size=64 * 1024 * 1024)
                     nbytes, rmse, sigma = _stream_quantize(
                         fin, fout, ti, abs_offset, 'q2k', fb_perms,
                         imatrix_data, use_hpc,
-                        fold_interleave=fold_interleave, perm_file=perm_file)
+                        fold_interleave=fold_interleave)
                     if rmse is not None:
                         q2k_rmse_sum += rmse
                         q2k_tensor_count += 1
@@ -1785,17 +1758,6 @@ def main():
                     quant_count += 1
                     total_quant_bytes += nbytes
                     pad = align_offset(fout.tell()) - fout.tell()
-                    if pad > 0:
-                        fout.write(b'\x00' * pad)
-                    if perm_file is not None:
-                        perm_file.seek(0)
-                        while True:
-                            chunk = perm_file.read(16 * 1024 * 1024)
-                            if not chunk:
-                                break
-                            fout.write(chunk)
-                        perm_file.close()
-                    pad = align_offset(fout.tell()) - fout.tell()
                     if pad > 0:
                         fout.write(b'\x00' * pad)
                     continue