krsnalyst commited on
Commit
a236231
·
verified ·
1 Parent(s): f5e7361

Archive unused LiteRT artifact: hf_archive_2026-06-14_README.md

Browse files
Files changed (1) hide show
  1. archive/2026-06-14/README.md +95 -0
archive/2026-06-14/README.md ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Pocket Talk LiteRT Model Archive - 2026-06-14
2
+
3
+ This archive contains local model artifacts that are no longer on the active Pocket Talk test path.
4
+
5
+ ## Current Artifacts Kept Locally
6
+
7
+ These are intentionally not part of this archive:
8
+
9
+ - `gemma-4-E2B-it.litertlm`
10
+ - Official Google/LiteRT community Gemma 4 E2B IT baseline.
11
+ - Known-good Pocket Talk reference model.
12
+ - Initializes and runs on the Moto G71 GPU at 2000+ tokens.
13
+ - Kept locally because it is the comparison baseline for every future model.
14
+
15
+ - `output/model-trevor-dyn-wi4-afp32-vision280.litertlm`
16
+ - Current best working TrevorJS abliterated Gemma 4 E2B candidate.
17
+ - Built via the community/Kaggle-style Gemma 4 LiteRT path, then spliced with official non-language sections.
18
+ - Patched so vision encoder/adapter expose only the `vision_280` signatures.
19
+ - Tested on phone: initializes on GPU at 2000 tokens, generates text, accepts images, and produces image-grounded responses.
20
+ - Kept locally because this is the active model we should keep improving.
21
+
22
+ ## Archived Model Artifacts
23
+
24
+ ### `peppx_gemma4_e2b_uncensored_MAX.litertlm`
25
+
26
+ - Local size: 2,550,041,824 bytes.
27
+ - SHA256: `34738F7EB50D04EE123D23DCEA0DF0B49E074B555134101251685F9416F39BB9`.
28
+ - Source role: PeppX community comparison candidate for an uncensored Gemma 4 E2B LiteRT-LM package.
29
+ - Structure noted during research:
30
+ - 5 sections only.
31
+ - Text-only public-export style bundle.
32
+ - Signatures include `prefill_128`, `prefill_256`, `prefill_512`, `prefill_1024`, and `decode`.
33
+ - KV cache and mask were FLOAT32.
34
+ - Target-path dequant count was 0.
35
+ - Why archived:
36
+ - Useful as a reference/community comparison, but not our current winning artifact.
37
+ - The current working model is the official-spliced Trevor WI4 Vision280 build, which has phone GPU and image success.
38
+
39
+ ### `gemma4_final.litertlm`
40
+
41
+ - Local size: 277,053,440 bytes.
42
+ - SHA256: `F0897EF21CDD3C86F364DE7FEA4D5931CBB82D34A159867654544787C784E7F4`.
43
+ - Source role: early local Gemma 4 LiteRT conversion attempt.
44
+ - Why archived:
45
+ - Much smaller than the real official Gemma 4 E2B LiteRT-LM package, so it is not comparable to the working 2.58 GB baseline.
46
+ - Superseded by later official-prefill, targeted-quantization, community recipe, and final Vision280 experiments.
47
+ - Not part of the current app allowlist or phone test path.
48
+
49
+ ### `model_gemma4_final.litertlm`
50
+
51
+ - Local size: 47,495,220 bytes.
52
+ - SHA256: `F6A5CEC0BA11263F4E0621F46650373A54C1C7952F1D56B0F2E14DDF18DEFA6B`.
53
+ - Source role: very early small Gemma 4 LiteRT artifact.
54
+ - Why archived:
55
+ - Too small to be a complete Gemma 4 E2B LiteRT-LM package.
56
+ - Kept only as historical evidence from the first conversion pass.
57
+ - Superseded by the working Trevor WI4 Vision280 model.
58
+
59
+ ## Archived Diagnostic Fragments
60
+
61
+ These `.bin` files are not runnable models. They are binary slices captured during debugging to compare official vs custom package internals.
62
+
63
+ - `official_head_16mb.bin`
64
+ - SHA256: `699C6EE24CF148758801933DD12D87ED3098C15E52F2868CB40C561DC056D08B`.
65
+ - First 16 MiB of the official baseline package.
66
+
67
+ - `official_prefill_head_8mb.bin`
68
+ - SHA256: `C1B3BEB79B275B510333A2B88C7D3199D4E9FBA15C83CAC2A1D0F0C03DA5313A`.
69
+ - First 8 MiB of the official prefill/decode section.
70
+
71
+ - `official_prefill_head_64mb.bin`
72
+ - SHA256: `F15645732C97911D0BABF0D18EFEC65117E002D61401F04CF0F0F250E9A47E8A`.
73
+ - First 64 MiB of the official prefill/decode section.
74
+
75
+ - `official_prefill_tail_from_800mb.bin`
76
+ - SHA256: `98F504DAACBF841B1BE18C61B73C8D75761C4ED3E0A1613C51786A968C01C53F`.
77
+ - Tail slice captured from around the official prefill/decode section.
78
+
79
+ - `modelihope_prefill_head_8mb.bin`
80
+ - SHA256: `BB914C6C0B9C31E0572A360AEC42FD15A7EB8AE87C829BB9D322FB8E2724E23D`.
81
+ - First 8 MiB of the early `modelihope` custom prefill/decode section.
82
+
83
+ - `modelthelasthope_prefill_head_12mb.bin`
84
+ - SHA256: `A3D9912B814E137D30BE5FD99AE469BCB22E3113D3C2357CA7F8CD65C099CB12`.
85
+ - First 12 MiB of the early `modelthelasthope` custom prefill/decode section.
86
+
87
+ ## Main Learning
88
+
89
+ The old failed path was not simply "model too large." The important pattern was model structure:
90
+
91
+ - Official Gemma worked because its runtime sections, signatures, and GPU-compatible graph layout matched LiteRT-LM expectations.
92
+ - Older direct public exports commonly exposed FLOAT32 KV/cache-related paths and produced Adreno GPU delegate shape mismatches.
93
+ - CPU fallback helped prove graph compatibility for some builds, but large custom models could still trigger Android low-memory kills.
94
+ - The working path was the community Gemma 4 export recipe plus official-section splicing, then patching the vision signatures down to `vision_280`.
95
+