OpenTransformer commited on
Commit
e47d112
·
verified ·
1 Parent(s): 02a452b

Add AGILLM-4 dblock single-file trainer

Browse files
README.md ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: pytorch
3
+ tags:
4
+ - pytorch
5
+ - transformer
6
+ - language-model
7
+ - long-context
8
+ - agillm
9
+ - dblock
10
+ - single-file
11
+ - experimental
12
+ ---
13
+
14
+ # AGILLM-4 dblock single-file
15
+
16
+ This repo packages the live AGILLM-4 dblock trainer as one runnable Python file:
17
+
18
+ - `agillm4_dblock_single_file.py`
19
+
20
+ It was generated on `2026-05-31T14:54:55Z` by mechanically inlining the live VastAI training sources:
21
+
22
+ - `fused_ce.py`
23
+ - `anchor_memory.py`
24
+ - `dblocks_train.py`
25
+ - `nB300_agillm4.py`
26
+
27
+ The original live command used `nB300_agillm4.py train`. This single-file build keeps that CLI surface, but registers in-memory shims for the former helper modules so the dblock path can still import `dblocks_train`, `anchor_memory`, and `fused_ce` internally.
28
+
29
+ See `single_file_manifest.json` for source hashes from the generated build.
30
+
31
+ Example shape:
32
+
33
+ ```bash
34
+ python agillm4_dblock_single_file.py train --preset agillm4_floor --dblock ...
35
+ ```
36
+
37
+ This is experimental training code, not a polished inference package.
agillm4_dblock_single_file.py ADDED
The diff for this file is too large to render. See raw diff
 
single_file_manifest.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "generated_at_utc": "2026-05-31T14:54:55Z",
3
+ "source_dir": "/workspace/agillm-4",
4
+ "source_files": [
5
+ {
6
+ "path": "fused_ce.py",
7
+ "bytes": 2435,
8
+ "sha256": "392d9ccc46cf5e750841a10d8244da41c42cdf93930c889f7027b7de74f3fd9d"
9
+ },
10
+ {
11
+ "path": "anchor_memory.py",
12
+ "bytes": 3114,
13
+ "sha256": "cfb41574628fe1a452dc29ab2c9c4147c5f4d128724306f25e229483939c558a"
14
+ },
15
+ {
16
+ "path": "dblocks_train.py",
17
+ "bytes": 16929,
18
+ "sha256": "a28b7d733c3c4b151dd5af2f4b1d3f3665da4efb424cbfd6d89ff012dcbe51cd"
19
+ },
20
+ {
21
+ "path": "nB300_agillm4.py",
22
+ "bytes": 147310,
23
+ "sha256": "6dd461d970364483ec0df6825b7177765dacda2115a1517e21ff53b51a34ca11"
24
+ }
25
+ ],
26
+ "entrypoint": "agillm4_dblock_single_file.py",
27
+ "repo_purpose": "Single-file AGILLM-4 dblock trainer packaging"
28
+ }