Add qwen3vl_strip/f2

qwen3vl_strip/f2/README.md

@@ -0,0 +1,36 @@
+# Qwen3-VL 8B line reader: strip, fold 2
+
+Adapter package for [`unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit`](https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit) at revision `b5b904c3fcdc7541adf2a2bb219b0ed95288c794`.
+Role `qwen3vl_strip`, fold 2 of 5. Input: one image of the whole line (height 160, width up to 5120).
+
+| Item | Value |
+|---|---|
+| Format | PEFT LoRA adapter, safetensors: `adapter_model.safetensors` (744 tensors) and `adapter_config.json` |
+| LoRA | rank 32, alpha 32, no dropout, no bias, on the attention and MLP projections of the vision tower and of the language model (720 float32 tensors) |
+| Complete modules | the visual merger and the three deepstack mergers (24 bfloat16 tensors) |
+| Base | `unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit` @ `b5b904c3fcdc7541adf2a2bb219b0ed95288c794`: Qwen3-VL-8B-Instruct with most language-model linear layers in 4-bit NF4 (bitsandbytes, double quantisation, bfloat16 compute); vision tower, mergers, embeddings and output head in bfloat16. Licence Apache-2.0 |
+| Processor | the base's processor, tokenizer and chat template, unchanged |
+
+## Loading
+
+No merge and no conversion. `models/qwen3vl/model.py` loads the base with Unsloth's
+`FastVisionModel.from_pretrained(..., load_in_4bit=True)`, builds the same LoRA layout with
+`FastVisionModel.get_peft_model` and loads every tensor of `adapter_model.safetensors` into it,
+checking each one bit for bit. The base is downloaded from its own repository at the revision
+above; it is not part of this package.
+
+## Decoding
+
+As set in `configs/recipes/qwen3vl_strip.yaml`: each decode process first runs one forward pass without
+gradient on four generated noise lines (recipe key `warmup`; it fixes torch.compile's state and
+changes no weight), then beam search with 4 beams over batches of 4 rows, at most 192 new tokens,
+length penalty 1, no sampling. A row that uses all 192 new tokens has a repeated
+tail cut by the cycle guard; when the guard cannot shorten it, the row is decoded again greedily
+in a fresh process, and if that also uses all 192 tokens the reader gives no answer for the row.
+
+## Files
+
+| File | Bytes | SHA-256 |
+|---|---:|---|
+| `adapter_model.safetensors` | 731,878,144 | `b4c17217cce4339f47b5db11e6014214583331afc04c3118a009ef545b9b2b2c` |
+| `adapter_config.json` | 1,838 | `069aae4c4abd2f5a1d391cd883287ca25ec0514993f7dd3268b3dffb5bd92005` |

qwen3vl_strip/f2/adapter_config.json

@@ -0,0 +1,51 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": {
+    "base_model_class": "Qwen3VLForConditionalGeneration",
+    "parent_library": "transformers.models.qwen3_vl.modeling_qwen3_vl",
+    "unsloth_fixed": true
+  },
+  "base_model_name_or_path": "unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": [
+    "model.visual.merger",
+    "model.visual.deepstack_merger_list.0",
+    "model.visual.deepstack_merger_list.1",
+    "model.visual.deepstack_merger_list.2"
+  ],
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.20.0",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": "b5b904c3fcdc7541adf2a2bb219b0ed95288c794",
+  "target_modules": "(?:.*?(?:vision|image|visual|patch|language|text).*?(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense|mixer).*?(?:qkv|proj|linear_fc1|linear_fc2|q_proj|k_proj|v_proj|o_proj|gate_proj|up_proj|down_proj))|(?:\\bmodel\\.layers\\.[\\d]{1,}\\.(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense|mixer)\\.(?:(?:qkv|proj|linear_fc1|linear_fc2|q_proj|k_proj|v_proj|o_proj|gate_proj|up_proj|down_proj)))",
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

qwen3vl_strip/f2/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4c17217cce4339f47b5db11e6014214583331afc04c3118a009ef545b9b2b2c
+size 731878144