litertlm_manifest.json: known_issues now carry the runtime notes added to the card on 2026-10-12/13 — one Engine per conversation (LiteRT-LM #3165); tool calls (LiteRT-LM #3927); 8 GB-class phones and Mali; the CPU int8 path

litertlm_manifest.json

@@ -1,7 +1,7 @@
 {
   "manifest_schema": "0.1.3",
   "repo": "mlboydaisuke/OvisOCR2-LiteRT",
-  "generated": "2026-10-07",
+  "generated": "2026-10-13",
   "generator": "make_manifest.py",
   "model": {
     "display_name": "OvisOCR2",
@@ -181,7 +181,11 @@
       ],
       "known_issues": [
         "General-chat sanity gates score low (Mac 8Q: 3-4/8) with a transcription/echo failure mode - checkpoint behavior from the OCR post-training, not quantization damage (anchored against the HF fp32 original)",
-        "Dense academic pages at 512x512 transcribe correctly at the start then drift into repetition; upstream's own reference code ships a repeat-cleanup post-processor for this"
+        "Dense academic pages at 512x512 transcribe correctly at the start then drift into repetition; upstream's own reference code ships a repeat-cleanup post-processor for this",
+        "Create a new Engine for each conversation: on litert-lm 0.18.0 a new Conversation on an Engine that already ran another conversation starts from the earlier conversation's recurrent state (LiteRT-LM #3165), and in a three-question probe (outside this model's task: it is a document parser that tends to transcribe a prompt instead of answering it) the _int8 file's second-conversation replies on the same Engine repeated 'The capital of France?' until the 200-token cap in both runs, while on a fresh Engine they began 'Paris', '17' and 'The quick brown fox jumps over the lazy dog.' and then repeated the fox sentence until the cap; the first conversation on an Engine is unaffected.",
+        "LiteRT-LM's CPU path quantizes the input of every int8 linear layer to int8 at run time (XNNPACK dynamic quantization); this file keeps the decoder recipe of the Qwen3.5-0.8B-VL_int8.litertlm file in litert-community/Qwen3.5-0.8B published until 2026-10-12 (the delta-rule decay/gate projections int8; that repo re-quantized its files that day), and the CPU transcription checks on the card (exact on the table page, where the CPU and GPU backends also produce byte-identical output) are the check on that path; a second check asks 'What is 17 times 23?' because the sibling Qwen3.5-2B int8 file looped on that question on the CPU path until its 2026-10-11 re-quantization, and in 3 fresh CPU processes the _int8 file repeated text until the 400-token cap each time without giving 391 (litert-lm 0.18.0, Mac CPU), a question outside this model's task (on its sanity gate the model echoes or loops on such prompts, and so does its fp32 original on the stock Hugging Face stack).",
+        "On Android this file's GPU engine has been built and run on a Galaxy S26 (Adreno, MemTotal 11,389,756 KiB) only (the transcription itself was not checked on that phone) and no Mali phone has run it; the Qwen3.5-0.8B-VL_int8.litertlm file in litert-community/Qwen3.5-0.8B published until 2026-10-12 shares this file's decoder recipe and six-signature prefill ladder, and on an 8 GB Pixel 8a (Arm Mali, MemTotal 7,754,700 KiB) that file's GPU engine creation reached 4 of the 6 prefill signatures before free memory fell below 400,000 KiB, the test's stop line (reading 300,460 KiB), so this file's GPU engine is not expected to fit an 8 GB-class phone (an expectation carried over from that file, not a measurement of this one): use the CPU backend on 8 GB phones.",
+        "This bundle's chat template has no tools section, so tool definitions never reach the prompt, and litert-lm 0.18.0 parses no Qwen3.5 tool call (such a call comes back as plain text and tool_calls stays empty; LiteRT-LM #3927): in a test of this file with a tool supplied and the runtime's automatic tool calling turned on, no tool ran (litert-lm 0.18.0 on a Mac)."
       ]
     }
   ]