# Qwen3-VL-8B-sft-mr-v1.1 / checkpoint-7199 -> HuggingFace format

Source: `/mnt/beegfs/checkpoints/xieqingsong/Qwen3-VL-8B-sft-mr-v1.1/checkpoint-7199`
Converted: 2026-08-28, verified against `transformers 4.57.6`.

The source is a **transformers 5.12.1** Trainer checkpoint. DeepSpeed had already
written a consolidated `model.safetensors` next to the ZeRO shards, so
`zero_to_fp32.py` was NOT needed — `global_step7199/` holds optimizer state only
(100 GB, training-resume material, not copied).

## Weights: untouched

All 750 tensors copied verbatim and verified **bit-identical** to the source
(`torch.equal` per tensor), all `bfloat16`, no NaN/Inf. Key names already match
the 4.57 layout (`model.language_model.*`, `model.visual.*`, `lm_head.weight`) —
no renaming was required.

## config.json: the one change that actually mattered

transformers 5.x renamed `text_config.rope_scaling` -> `rope_parameters` and moved
`rope_theta` inside it. **4.57 looks for `rope_scaling` only.** Loading the source
checkpoint under 4.57 therefore succeeds with no error while silently dropping
`mrope_section: [24, 20, 20]` and `mrope_interleaved: true`:

    AutoConfig.from_pretrained(checkpoint-7199).text_config.rope_scaling  ->  None

i.e. the model falls back to plain RoPE and every image gets the wrong positional
encoding, with nothing in the logs to say so. The conversion restores the 4.57
layout; the converted dir asserts `mrope_section == [24, 20, 20]` on load.

Other config edits, all schema-only:

| key | source (tf 5.12.1) | converted (tf 4.57) | why |
|---|---|---|---|
| `text_config.rope_parameters` | present | -> `rope_scaling` + `rope_theta` | see above |
| `vision_config.model_type` | `qwen3_vl_vision` | `qwen3_vl` | 4.57 registry name |
| `use_cache` (top level) | `false` | removed | training artifact; `text_config.use_cache` stays `true` |
| `bos/eos/pad_token_id`, `dtype` (top level) | present | removed | not in 4.57's Qwen3VLConfig; they live in `generation_config.json` |
| `text_config.pad_token_id` | `null` | removed | tf-5.x writes an explicit null |
| `vision_config.dtype` | `bfloat16` | removed | inferred from tensors |
| `transformers_version` | `5.12.1` | `4.57.0` | |

`generation_config.json`: `eos_token_id` deduped `[151645, 151645, 151643]` ->
`[151645, 151643]` (the duplicate was a tf-5.x save artifact), `bos_token_id:
151643` added to match the base model. Sampling params left as trained.

## Tokenizer / processor

Taken from `/mnt/beegfs/weights/Qwen3-VL-8B-Instruct` (the SFT base). This is
sound, and was checked rather than assumed:

* vocab (151643 entries) — **identical**
* added_tokens (26) — **identical**
* merges (151387) — identical once the tf-5.x list-of-pairs serialization
  (`["Ġ","Ġ"]`) is normalized to 4.x's string form (`"Ġ Ġ"`)
* chat template — **byte-identical** to the checkpoint's own `chat_template.jinja`

The SFT trained `embed_tokens` but added no tokens, so the vocabulary is the
base's. `processor_config.json` (tf-5.x nested format) was dropped in favour of
`preprocessor_config.json` + `video_preprocessor_config.json`; their values agree
(min_pixels 65536 / max_pixels 16777216, matching the training `DataArguments`).

## Not copied

`global_step7199/`, `rng_state_*.pth`, `scheduler.pt`, `trainer_state.json`,
`training_args.bin`, `latest`, `zero_to_fp32.py`, `tokenizer_config.json.tf5bak`.

## Verified

    AutoConfig / AutoProcessor / AutoModelForImageTextToText load under tf 4.57.6
    8.77B params, bf16, generate() on a test image returns coherent text.

Note `training_args.bin` cannot be unpickled outside the training env (it needs
the `qwenvl` module); the training meta was
`datav2/train_data_traj_v2_mr_spatial_stepinstr.json` (navigation + grounding
styles, conv template `internvl2_5_mobile_navigation_v1`).