CompressedGemma commited on
Commit
262cc7b
·
verified ·
1 Parent(s): dc3b370

Update code comments

Browse files
Files changed (1) hide show
  1. hexstate_requantize.py +17 -5
hexstate_requantize.py CHANGED
@@ -1,15 +1,27 @@
1
  #!/usr/bin/env python3
2
  """
3
- HExState GGUF Re-Quantizer — GGUF-to-GGUF Q2_K quantization.
4
 
5
  Reads a source GGUF (F16/BF16/F32), copies all metadata verbatim,
6
- and re-quantizes eligible weight tensors to Q2_K using numpy.
 
7
 
8
- This bypasses the tokenizer parsing problem entirely — the source GGUF
9
- (from llama.cpp's convert_hf_to_gguf.py) has correct metadata.
 
 
 
 
10
 
11
  Usage:
12
- python3 hexstate_requantize.py input.gguf output.gguf
 
 
 
 
 
 
 
13
  """
14
 
15
  import struct
 
1
  #!/usr/bin/env python3
2
  """
3
+ HexState GGUF Re-Quantizer — GGUF-to-GGUF HPC quantization.
4
 
5
  Reads a source GGUF (F16/BF16/F32), copies all metadata verbatim,
6
+ and re-quantizes eligible weight tensors using the HExState HPC engine
7
+ (Shor-optimized Griffiths-Niu measurement via libhexstate_q2k.so).
8
 
9
+ Quantization tiers:
10
+ - Attention Q/K/V/O + DeltaNet SSM projections → Q4_0 (HPC-optimized)
11
+ - FFN / MLP weight matrices → Q2_K (HPC-optimized)
12
+ - Embeddings, norms, biases, LM head → kept at source precision
13
+
14
+ Falls back to a pure numpy Q2_K implementation if the C library is not built.
15
 
16
  Usage:
17
+ python3 hexstate_requantize.py input.gguf output.gguf [options]
18
+
19
+ Options:
20
+ --config <file> Load HuggingFace config.json for arch detection
21
+ --imatrix <file> Importance matrix for calibrated quantization
22
+ --keep-metadata Preserve all GGUF metadata as-is
23
+ --q2all Force all eligible tensors to Q2_K
24
+ --quantize-none Skip quantization (passthrough)
25
  """
26
 
27
  import struct