OpenTransformer commited on
Commit
bbcc751
·
verified ·
1 Parent(s): d4ce194

Add AGILLM-4.3 inference sample at step 2161945

Browse files

Adds prompt/output smoke test for checkpoint final.pt step 2161945, plus raw result artifact.

README.md CHANGED
@@ -75,3 +75,30 @@ If you want to host your own decentralized training swarm for AGILLM models, you
75
  --base-ckpt /path/to/base_model.pt
76
  ```
77
  Validated updates will be moved to an `accepted/` directory which your master can then asynchronously merge.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
75
  --base-ckpt /path/to/base_model.pt
76
  ```
77
  Validated updates will be moved to an `accepted/` directory which your master can then asynchronously merge.
78
+
79
+ <!-- agillm43-latest-inference-start -->
80
+ ## Latest Inference Smoke Test
81
+
82
+ This is a raw quality smoke test from the active AGILLM-4.3 training checkpoint. It is included to show that the checkpoint loads and decodes; the text is not yet meant to be good instruction-following output.
83
+
84
+ - Created UTC: `2026-06-18T05:13:28Z`
85
+ - Checkpoint: `final.pt`
86
+ - Checkpoint step: `2161945`
87
+ - Method: `ar_block_stream_fp16_sdpa_ignore_eos`
88
+ - Decode stats: `55.85s | 64 tokens | 1.1 tok/s`
89
+ - Full artifact: [`inference_samples/step2161945_20260618T051328Z/README.md`](./inference_samples/step2161945_20260618T051328Z/README.md)
90
+
91
+ ### Prompt
92
+
93
+ ```text
94
+ In one short paragraph, explain what AGILLM4.3 is doing during this training run.
95
+ ```
96
+
97
+ ### Output
98
+
99
+ ```text
100
+ If you can be a lot of the same time to do that we are not like it's I'm so if they have some things and you don't know about your way for my book.
101
+
102
+ I think how you want to get me! You're just in our thing, I've never had been on the other
103
+ ```
104
+ <!-- agillm43-latest-inference-end -->
inference_samples/step2161945_20260618T051328Z/README.md ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AGILLM-4.3 inference sample
2
+
3
+ - Created UTC: `2026-06-18T05:13:28Z`
4
+ - Checkpoint: `final.pt`
5
+ - Checkpoint path on training box: `/workspace/agillm4_4090_ckpts_active/final.pt`
6
+ - Checkpoint step: `2161945`
7
+ - Method: `ar_block_stream_fp16_sdpa_ignore_eos`
8
+ - Decode stats: `55.85s | 64 tokens | 1.1 tok/s`
9
+ - Note: this smoke test used `ignore_eos` to force a visible base-model continuation after the same checkpoint otherwise emitted EOS immediately.
10
+
11
+ ## Prompt
12
+
13
+ ```text
14
+ In one short paragraph, explain what AGILLM4.3 is doing during this training run.
15
+ ```
16
+
17
+ ## Output
18
+
19
+ ```text
20
+ If you can be a lot of the same time to do that we are not like it's I'm so if they have some things and you don't know about your way for my book.
21
+
22
+ I think how you want to get me! You're just in our thing, I've never had been on the other
23
+ ```
inference_samples/step2161945_20260618T051328Z/infer_ar_ignore_eos.txt ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [tokenizer] transformers=4.48.0, tokenizers=0.21.4
2
+ [tokenizer] Health check PASSED - spaces preserved in roundtrip
3
+ [tokenizer] Restored from /workspace/agillm4_4090_ckpts_active/final.pt.tokenizer.json
4
+ [infer] inference_time=2026-06-18 06:12:17 BST
5
+ [infer] checkpoint=final.pt
6
+ [attention_config]
7
+ d_model=1280 heads=20 d_k=64
8
+ layers=28 tie_weights=YES
9
+ rank=160 ratio=2.5x regime=EXPANSION
10
+ [infer] block_stream_dtype=float16
11
+ [infer] block_stream=True device=cuda page=1 layer(s); page_cache=False (explicit-off) KV cache=True kv_device=cpu
12
+ Model size: 1.22B parameters (1,221,580,802)
13
+ Generating (ar)...
14
+ [infer] block_stream_cuda_peak_alloc=0.42GB peak_reserved=0.44GB
15
+ In one short paragraph, explain what AGILLM4.3 is doing during this training run. If you can be a lot of the same time to do that we are not like it's I'm so if they have some things and you don't know about your way for my book.
16
+
17
+ I think how you want to get me! You're just in our thing, I've never had been on the other
18
+ [55.85s | 64 tokens | 1.1 tok/s]
19
+ [CLAUDE_FRIENDLY_START]
20
+ [mode=ar]
21
+ [prompt_input]
22
+ In one short paragraph, explain what AGILLM4.3 is doing during this training run.
23
+ [completion]
24
+ If you can be a lot of the same time to do that we are not like it's I'm so if they have some things and you don't know about your way for my book.
25
+
26
+ I think how you want to get me! You're just in our thing, I've never had been on the other
27
+ [prompt_plus_completion]
28
+ In one short paragraph, explain what AGILLM4.3 is doing during this training run. If you can be a lot of the same time to do that we are not like it's I'm so if they have some things and you don't know about your way for my book.
29
+
30
+ I think how you want to get me! You're just in our thing, I've never had been on the other
31
+ [stats] 55.85s | 64 tokens | 1.1 tok/s
32
+ [CLAUDE_FRIENDLY_END]
33
+ [wall_time] 80s
inference_samples/step2161945_20260618T051328Z/metadata.txt ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ AGILLM43_INFER_START 2026-06-18T05:12:08Z
2
+ checkpoint=/workspace/agillm4_4090_ckpts_active/final.pt
3
+ prompt=In one short paragraph, explain what AGILLM4.3 is doing during this training run.
4
+ mode=ar sampler=ar max_new=64 min_new=64 ignore_eos greedy block_stream fp16 sdpa
inference_samples/step2161945_20260618T051328Z/result.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_at": "2026-06-18T05:13:28Z",
3
+ "checkpoint": "/workspace/agillm4_4090_ckpts_active/final.pt",
4
+ "checkpoint_latest_json": "{\"path\": \"/workspace/agillm4_4090_ckpts_active/final.pt\", \"step\": 2161945}",
5
+ "method": "ar_block_stream_fp16_sdpa_ignore_eos",
6
+ "prompt": "In one short paragraph, explain what AGILLM4.3 is doing during this training run.",
7
+ "completion": "If you can be a lot of the same time to do that we are not like it's I'm so if they have some things and you don't know about your way for my book.\n\nI think how you want to get me! You're just in our thing, I've never had been on the other",
8
+ "stats": "55.85s | 64 tokens | 1.1 tok/s",
9
+ "ok": true,
10
+ "raw_file": "infer_ar_ignore_eos.txt",
11
+ "checkpoint_step": 2161945,
12
+ "checkpoint_name": "final.pt",
13
+ "checkpoint_path": "/workspace/agillm4_4090_ckpts_active/final.pt",
14
+ "huggingface_repo": "OpenTransformer/AGILLM-4.3"
15
+ }
inference_samples/step2161945_20260618T051328Z/run_README.md ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AGILLM4.3 Inference Smoke Test
2
+
3
+ - Created: `2026-06-18T05:13:28Z`
4
+ - Checkpoint: `final.pt`
5
+ - Method: `ar_block_stream_fp16_sdpa_ignore_eos`
6
+ - OK: `True`
7
+
8
+ ## Prompt
9
+
10
+ In one short paragraph, explain what AGILLM4.3 is doing during this training run.
11
+
12
+ ## Completion
13
+
14
+ ```text
15
+ If you can be a lot of the same time to do that we are not like it's I'm so if they have some things and you don't know about your way for my book.
16
+
17
+ I think how you want to get me! You're just in our thing, I've never had been on the other
18
+ ```
19
+
20
+ ## Stats
21
+
22
+ `55.85s | 64 tokens | 1.1 tok/s`