Add AWQ W4A16 g128 checkpoint, model card, and quantization metadata

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

NOTICE

@@ -0,0 +1,13 @@
+InternVL3.5-4B-HF AWQ W4A16 (group size 128)
+
+This model is a quantized derivative of:
+OpenGVLab/InternVL3_5-4B-HF
+https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF
+
+The upstream project is licensed under the Apache License 2.0.
+The quantized checkpoint preserves the upstream model architecture and files,
+with the language decoder Linear weights compressed to asymmetric INT4
+(group size 128) using the AWQ algorithm via llm-compressor.
+
+The vision tower, multimodal projector, input embeddings, and lm_head are
+left unquantized in BF16.

QUANTIZATION_INFO.json

@@ -0,0 +1,42 @@
+{
+  "source_directory": "/home/bellock/projects/internvl35-fp8/models/InternVL3_5-4B-HF",
+  "output_directory": "/home/bellock/projects/internvl35-fp8/models/InternVL3_5-4B-AWQ-W4A16-G128",
+  "source_revision": "model_id=OpenGVLab/InternVL3_5-4B-HF\nrevision=6bd4487402110ef9889ba50eb7aefeb302526fed",
+  "quantization_scheme": "W4A16_ASYM",
+  "quantization_algorithm": "AWQ",
+  "weight_format": "pack-quantized",
+  "weight_num_bits": 4,
+  "weight_type": "int",
+  "weight_strategy": "group",
+  "weight_group_size": 128,
+  "weight_symmetric": false,
+  "input_activations": null,
+  "target_module_type": "Linear (language decoder only, regex-scoped)",
+  "target_linear_count": 252,
+  "target_attention_linear": 144,
+  "target_mlp_linear": 108,
+  "protected_linear_count": 147,
+  "protected_modules": [
+    "vision_tower",
+    "multi_modal_projector",
+    "lm_head",
+    "embed_tokens"
+  ],
+  "awq_duo_scaling": true,
+  "awq_n_grid": 20,
+  "calibration_dataset": "lmms-lab/flickr30k",
+  "calibration_samples": 128,
+  "calibration_skipped": 0,
+  "calibration_image_size": "448x448",
+  "calibration_image_patches_per_sample": 1,
+  "calibration_sequence_length_range": "278-288",
+  "source_dtype": "torch.bfloat16",
+  "python_version": "3.12.3",
+  "torch_version": "2.12.0+cu132",
+  "transformers_version": "5.10.1",
+  "llmcompressor_version": "0.12.0.1",
+  "compressed_tensors_version": "0.17.1",
+  "checkpoint_tensors_total": 1597,
+  "checkpoint_packed_int4_modules": 252,
+  "checkpoint_size_gib": 3.816
+}

README.md

@@ -0,0 +1,227 @@
+---
+license: apache-2.0
+library_name: transformers
+pipeline_tag: image-text-to-text
+base_model: OpenGVLab/InternVL3_5-4B-HF
+base_model_relation: quantized
+language:
+  - en
+  - zh
+  - ko
+tags:
+  - internvl
+  - internvl3.5
+  - vision-language
+  - multimodal
+  - vllm
+  - compressed-tensors
+  - awq
+  - int4
+  - w4a16
+  - ampere
+  - wsl2
+---
+
+# InternVL3.5-4B-HF AWQ W4A16 (group size 128)
+
+A compressed-tensors **AWQ W4A16** quantization of
+[OpenGVLab/InternVL3_5-4B-HF](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF),
+built for serving a 4B vision-language model on an 8 GB consumer GPU with vLLM.
+
+A companion FP8 build of the same base model is at
+[hsmin92/internvl35-fp8](https://huggingface.co/hsmin92/internvl35-fp8).
+See [Choosing between the FP8 and AWQ builds](#choosing-between-the-fp8-and-awq-builds).
+
+## Quantization scope
+
+Quantized to asymmetric INT4:
+
+- **252** language decoder `Linear` modules
+  (144 attention: `q_proj` / `k_proj` / `v_proj` / `o_proj`,
+  108 MLP: `gate_proj` / `up_proj` / `down_proj`)
+- Scheme: `W4A16_ASYM`, algorithm: AWQ (`duo_scaling`, `n_grid=20`)
+- Weight format: `pack-quantized`, group size **128**
+- Activations: **not** quantized (A16)
+
+Kept in BF16:
+
+- Vision tower (`vision_tower`)
+- Multimodal projector (`multi_modal_projector`)
+- Input embeddings
+- `lm_head`
+- Normalization layers and other protected parameters
+
+Generated from the base-model revision:
+
+```text
+6bd4487402110ef9889ba50eb7aefeb302526fed
+```
+
+Checkpoint contents verified: 1,597 tensors, 252 packed INT4 modules with
+matching scale / zero-point tensors, no packed weights under `vision_tower`,
+`multi_modal_projector`, `lm_head`, or the embeddings, and no NaN/Inf in any
+scale. See [`quantization/recipe.py`](./quantization/recipe.py).
+
+## Calibration
+
+AWQ fits per-channel scales against a calibration set, so the calibration
+distribution matters. This build used:
+
+| | |
+|---|---|
+| Dataset | `lmms-lab/flickr30k` |
+| Samples | 128 (0 skipped) |
+| Image size | 448×448, 1 patch per sample |
+| Sequence length | 278–288 tokens |
+| Prompts | generic English / Korean scene-description instructions |
+
+**This is a general-purpose photo set.** If your target domain is far from
+everyday photography — industrial inspection, thermal imagery, medical,
+document OCR, CCTV at unusual angles — re-run AWQ with in-domain calibration
+images rather than assuming this checkpoint transfers. The FP8 build needs no
+calibration and does not carry this caveat.
+
+## Verified environment
+
+| Component | Version / value |
+|---|---|
+| GPU | NVIDIA GeForce RTX 3070 8 GB (Ampere, SM 8.6) |
+| Host | Windows 11 + WSL2 (Ubuntu 24.04), Docker Desktop |
+| NVIDIA driver | 591.86 |
+| Serving image | `vllm/vllm-openai:v0.26.0` |
+| vLLM | 0.26.0 (V1 engine) |
+| PyTorch | 2.11.0+cu130 |
+| Transformers | 5.14.1 |
+| Quantization backend | compressed-tensors 0.17.0 |
+| Attention backend | FlashAttention (auto-selected) |
+
+Measured at startup with `--gpu-memory-utilization 0.86 --max-model-len 8192
+--max-num-seqs 1 --max-num-batched-tokens 4096 --dtype bfloat16`:
+
+```text
+Model loading took 3.84 GiB memory
+Available KV cache memory: 2.66 GiB
+GPU KV cache size: 19,344 tokens
+Maximum concurrency for 8,192 tokens per request: 2.36x
+Graph capturing finished in 1 secs, took 0.00 GiB
+```
+
+CUDA graphs capture successfully and full FP16 KV cache fits at an 8,192-token
+context — neither is possible with the FP8 build on the same 8 GB card.
+
+## Serving with vLLM
+
+`--dtype bfloat16` is required; see [Do not use `--dtype half`](#do-not-use---dtype-half).
+
+```bash
+vllm serve hsmin92/internvl35-4b-awq-w4a16-g128 \
+  --served-model-name internvl35-awq \
+  --dtype bfloat16 \
+  --max-model-len 8192 \
+  --max-num-seqs 1 \
+  --max-num-batched-tokens 4096 \
+  --limit-mm-per-prompt '{"image":2,"video":0}' \
+  --gpu-memory-utilization 0.86 \
+  --trust-remote-code
+```
+
+Then use the OpenAI-compatible endpoint at `http://127.0.0.1:8000/v1`.
+
+### Do not use `--dtype half`
+
+On this checkpoint family `--dtype half` produces a server that starts
+normally, returns HTTP 200 from `/health`, reports a healthy Docker
+healthcheck — and answers every request with `!!!!!!!!`. Dequantization
+overflows the FP16 range, the logits become NaN, and argmax selects token 0.
+It fails silently, so a liveness probe will not catch it.
+
+The checkpoint declares `bfloat16` in `config.json` and Ampere supports BF16
+natively at the same memory cost. Always send a real request after a
+configuration change, not just a health check.
+
+### `--max-num-batched-tokens` and multi-image requests
+
+This value also sizes the multimodal encoder cache. InternVL dynamic tiling
+allows up to 12 patches plus a thumbnail, so a single image can reach
+**3,329 embedding tokens**. With a smaller value, large images are rejected:
+
+```text
+image item with 2816 embedding tokens, which exceeds the
+pre-allocated encoder cache size 2048
+```
+
+Use at least 4096 if you send high-resolution or wide-aspect images.
+
+### Image token cost is driven by aspect ratio, not size
+
+Tile count is chosen from the aspect ratio, so token cost is not monotonic in
+resolution. Measured on this model:
+
+| Input size | Image tokens | Tiles |
+|---|---|---|
+| 448×448 | 257 | 1 |
+| 896×448 | 769 | 3 |
+| 960×544 | 769 | 3 |
+| 200×150 | **3,329** | **13** |
+| 800×450 | 769 | 3 |
+| 4000×300 | **3,329** | **13** |
+
+A 200×150 crop costs **4.3× more** than an 800×450 image, because 4:3 maps
+exactly onto a 4×3 tile grid and the crop is upscaled to 1792×1344 for no
+added information. When sending region crops, letterbox them into a fixed
+448×448 canvas (preserving aspect ratio, no upscaling) to make cost constant.
+
+## Choosing between the FP8 and AWQ builds
+
+Measured on the same RTX 3070, same vLLM version, same settings:
+
+| | FP8 Dynamic | **AWQ W4A16 G128** |
+|---|---|---|
+| Weights on GPU | 5.51 GiB | **3.84 GiB** |
+| KV cache available | 0.61 GiB | **2.66 GiB** |
+| KV cache capacity | 8,928 tokens | **19,344 tokens** |
+| Max concurrency @ 8192 ctx | 1.09× | **2.36×** |
+| KV cache dtype needed for 8192 ctx | fp8 (compromise) | **fp16** |
+| CUDA graphs on 8 GB | not possible | **yes** |
+| Text generation | 13–17 tok/s | **~55 tok/s** |
+| Calibration data required | **no** | yes (see above) |
+| Weight precision | 8-bit | 4-bit |
+
+AWQ is decisively better on memory and speed. **Accuracy has not been compared
+between the two builds, or against the BF16 base model.** 4-bit weights and
+out-of-domain calibration are both reasons to expect the AWQ build to degrade
+first on hard inputs. Benchmark both on your own data before choosing.
+
+## Limitations
+
+- A quantized derivative, not an independently trained model.
+- **No accuracy benchmark has been published for this checkpoint.** Functional
+  validation only: text generation, a 7k-token context, and multi-image
+  (full frame + crop) requests all produce coherent, correct answers on simple
+  synthetic probes.
+- Calibration is general-purpose photography; see [Calibration](#calibration).
+- The vision tower and output head remain BF16 and are a meaningful share of
+  the loaded weights, so the size reduction is smaller than 4/16 would suggest.
+- The bundled `chat_template.jinja` has no tool-calling support. Passing
+  `tools=[...]` is silently ignored regardless of vLLM's tool-parser flags.
+- VRAM figures depend on driver, desktop applications, context length,
+  multimodal limits, and vLLM version.
+
+## Attribution and license
+
+A quantized derivative of
+[OpenGVLab/InternVL3_5-4B-HF](https://huggingface.co/OpenGVLab/InternVL3_5-4B-HF).
+The original project and this derivative are distributed under the Apache-2.0
+license. Review the upstream model card for original training details,
+limitations, and citation information.
+
+## Citation
+
+```bibtex
+@article{wang2025internvl3_5,
+  title={InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency},
+  author={Wang, Weiyun and Gao, Zhangwei and Gu, Lixin and Pu, Hengjun and Cui, Long and Wei, Xingguang and Liu, Zhaoyang and Jing, Linglin and Ye, Shenglong and Shao, Jie and others},
+  journal={arXiv preprint arXiv:2508.18265},
+  year={2025}
+}
+```

chat_template.jinja

@@ -0,0 +1,6 @@
+{% for message in messages %}{{'<|im_start|>' + message['role'] + '
+'}}{% if message['content'] is string %}{{ message['content'] }}{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' %}{{ '<IMG_CONTEXT>
+' }}{% elif content['type'] == 'video' %}{{ '<video>
+' }}{% elif content['type'] == 'text' %}{{ content['text'] }}{% endif %}{% endfor %}{% endif %}{{'<|im_end|>
+'}}{% endfor %}{% if add_generation_prompt %}{{'<|im_start|>assistant
+' }}{% endif %}

config.json

@@ -0,0 +1,310 @@
+{
+  "architectures": [
+    "InternVLForConditionalGeneration"
+  ],
+  "downsample_ratio": 0.5,
+  "dtype": "bfloat16",
+  "image_seq_length": 256,
+  "image_token_id": 151671,
+  "model_type": "internvl",
+  "projector_hidden_act": "gelu",
+  "quantization_config": {
+    "config_groups": {
+      "group_0": {
+        "format": "pack-quantized",
+        "input_activations": null,
+        "output_activations": null,
+        "targets": [
+          "re:.*layers\\.\\d+\\.self_attn\\.(q_proj|k_proj|v_proj|o_proj)$",
+          "re:.*layers\\.\\d+\\.mlp\\.(gate_proj|up_proj|down_proj)$"
+        ],
+        "weights": {
+          "actorder": null,
+          "block_structure": null,
+          "dynamic": false,
+          "group_size": 128,
+          "num_bits": 4,
+          "observer": "memoryless_minmax",
+          "observer_kwargs": {},
+          "scale_dtype": null,
+          "strategy": "group",
+          "symmetric": false,
+          "type": "int",
+          "zp_dtype": "torch.int8"
+        }
+      }
+    },
+    "format": "pack-quantized",
+    "global_compression_ratio": null,
+    "ignore": [
+      "model.vision_tower.encoder.layer.0.attention.q_proj",
+      "model.vision_tower.encoder.layer.0.attention.k_proj",
+      "model.vision_tower.encoder.layer.0.attention.v_proj",
+      "model.vision_tower.encoder.layer.0.attention.projection_layer",
+      "model.vision_tower.encoder.layer.0.mlp.fc1",
+      "model.vision_tower.encoder.layer.0.mlp.fc2",
+      "model.vision_tower.encoder.layer.1.attention.q_proj",
+      "model.vision_tower.encoder.layer.1.attention.k_proj",
+      "model.vision_tower.encoder.layer.1.attention.v_proj",
+      "model.vision_tower.encoder.layer.1.attention.projection_layer",
+      "model.vision_tower.encoder.layer.1.mlp.fc1",
+      "model.vision_tower.encoder.layer.1.mlp.fc2",
+      "model.vision_tower.encoder.layer.2.attention.q_proj",
+      "model.vision_tower.encoder.layer.2.attention.k_proj",
+      "model.vision_tower.encoder.layer.2.attention.v_proj",
+      "model.vision_tower.encoder.layer.2.attention.projection_layer",
+      "model.vision_tower.encoder.layer.2.mlp.fc1",
+      "model.vision_tower.encoder.layer.2.mlp.fc2",
+      "model.vision_tower.encoder.layer.3.attention.q_proj",
+      "model.vision_tower.encoder.layer.3.attention.k_proj",
+      "model.vision_tower.encoder.layer.3.attention.v_proj",
+      "model.vision_tower.encoder.layer.3.attention.projection_layer",
+      "model.vision_tower.encoder.layer.3.mlp.fc1",
+      "model.vision_tower.encoder.layer.3.mlp.fc2",
+      "model.vision_tower.encoder.layer.4.attention.q_proj",
+      "model.vision_tower.encoder.layer.4.attention.k_proj",
+      "model.vision_tower.encoder.layer.4.attention.v_proj",
+      "model.vision_tower.encoder.layer.4.attention.projection_layer",
+      "model.vision_tower.encoder.layer.4.mlp.fc1",
+      "model.vision_tower.encoder.layer.4.mlp.fc2",
+      "model.vision_tower.encoder.layer.5.attention.q_proj",
+      "model.vision_tower.encoder.layer.5.attention.k_proj",
+      "model.vision_tower.encoder.layer.5.attention.v_proj",
+      "model.vision_tower.encoder.layer.5.attention.projection_layer",
+      "model.vision_tower.encoder.layer.5.mlp.fc1",
+      "model.vision_tower.encoder.layer.5.mlp.fc2",
+      "model.vision_tower.encoder.layer.6.attention.q_proj",
+      "model.vision_tower.encoder.layer.6.attention.k_proj",
+      "model.vision_tower.encoder.layer.6.attention.v_proj",
+      "model.vision_tower.encoder.layer.6.attention.projection_layer",
+      "model.vision_tower.encoder.layer.6.mlp.fc1",
+      "model.vision_tower.encoder.layer.6.mlp.fc2",
+      "model.vision_tower.encoder.layer.7.attention.q_proj",
+      "model.vision_tower.encoder.layer.7.attention.k_proj",
+      "model.vision_tower.encoder.layer.7.attention.v_proj",
+      "model.vision_tower.encoder.layer.7.attention.projection_layer",
+      "model.vision_tower.encoder.layer.7.mlp.fc1",
+      "model.vision_tower.encoder.layer.7.mlp.fc2",
+      "model.vision_tower.encoder.layer.8.attention.q_proj",
+      "model.vision_tower.encoder.layer.8.attention.k_proj",
+      "model.vision_tower.encoder.layer.8.attention.v_proj",
+      "model.vision_tower.encoder.layer.8.attention.projection_layer",
+      "model.vision_tower.encoder.layer.8.mlp.fc1",
+      "model.vision_tower.encoder.layer.8.mlp.fc2",
+      "model.vision_tower.encoder.layer.9.attention.q_proj",
+      "model.vision_tower.encoder.layer.9.attention.k_proj",
+      "model.vision_tower.encoder.layer.9.attention.v_proj",
+      "model.vision_tower.encoder.layer.9.attention.projection_layer",
+      "model.vision_tower.encoder.layer.9.mlp.fc1",
+      "model.vision_tower.encoder.layer.9.mlp.fc2",
+      "model.vision_tower.encoder.layer.10.attention.q_proj",
+      "model.vision_tower.encoder.layer.10.attention.k_proj",
+      "model.vision_tower.encoder.layer.10.attention.v_proj",
+      "model.vision_tower.encoder.layer.10.attention.projection_layer",
+      "model.vision_tower.encoder.layer.10.mlp.fc1",
+      "model.vision_tower.encoder.layer.10.mlp.fc2",
+      "model.vision_tower.encoder.layer.11.attention.q_proj",
+      "model.vision_tower.encoder.layer.11.attention.k_proj",
+      "model.vision_tower.encoder.layer.11.attention.v_proj",
+      "model.vision_tower.encoder.layer.11.attention.projection_layer",
+      "model.vision_tower.encoder.layer.11.mlp.fc1",
+      "model.vision_tower.encoder.layer.11.mlp.fc2",
+      "model.vision_tower.encoder.layer.12.attention.q_proj",
+      "model.vision_tower.encoder.layer.12.attention.k_proj",
+      "model.vision_tower.encoder.layer.12.attention.v_proj",
+      "model.vision_tower.encoder.layer.12.attention.projection_layer",
+      "model.vision_tower.encoder.layer.12.mlp.fc1",
+      "model.vision_tower.encoder.layer.12.mlp.fc2",
+      "model.vision_tower.encoder.layer.13.attention.q_proj",
+      "model.vision_tower.encoder.layer.13.attention.k_proj",
+      "model.vision_tower.encoder.layer.13.attention.v_proj",
+      "model.vision_tower.encoder.layer.13.attention.projection_layer",
+      "model.vision_tower.encoder.layer.13.mlp.fc1",
+      "model.vision_tower.encoder.layer.13.mlp.fc2",
+      "model.vision_tower.encoder.layer.14.attention.q_proj",
+      "model.vision_tower.encoder.layer.14.attention.k_proj",
+      "model.vision_tower.encoder.layer.14.attention.v_proj",
+      "model.vision_tower.encoder.layer.14.attention.projection_layer",
+      "model.vision_tower.encoder.layer.14.mlp.fc1",
+      "model.vision_tower.encoder.layer.14.mlp.fc2",
+      "model.vision_tower.encoder.layer.15.attention.q_proj",
+      "model.vision_tower.encoder.layer.15.attention.k_proj",
+      "model.vision_tower.encoder.layer.15.attention.v_proj",
+      "model.vision_tower.encoder.layer.15.attention.projection_layer",
+      "model.vision_tower.encoder.layer.15.mlp.fc1",
+      "model.vision_tower.encoder.layer.15.mlp.fc2",
+      "model.vision_tower.encoder.layer.16.attention.q_proj",
+      "model.vision_tower.encoder.layer.16.attention.k_proj",
+      "model.vision_tower.encoder.layer.16.attention.v_proj",
+      "model.vision_tower.encoder.layer.16.attention.projection_layer",
+      "model.vision_tower.encoder.layer.16.mlp.fc1",
+      "model.vision_tower.encoder.layer.16.mlp.fc2",
+      "model.vision_tower.encoder.layer.17.attention.q_proj",
+      "model.vision_tower.encoder.layer.17.attention.k_proj",
+      "model.vision_tower.encoder.layer.17.attention.v_proj",
+      "model.vision_tower.encoder.layer.17.attention.projection_layer",
+      "model.vision_tower.encoder.layer.17.mlp.fc1",
+      "model.vision_tower.encoder.layer.17.mlp.fc2",
+      "model.vision_tower.encoder.layer.18.attention.q_proj",
+      "model.vision_tower.encoder.layer.18.attention.k_proj",
+      "model.vision_tower.encoder.layer.18.attention.v_proj",
+      "model.vision_tower.encoder.layer.18.attention.projection_layer",
+      "model.vision_tower.encoder.layer.18.mlp.fc1",
+      "model.vision_tower.encoder.layer.18.mlp.fc2",
+      "model.vision_tower.encoder.layer.19.attention.q_proj",
+      "model.vision_tower.encoder.layer.19.attention.k_proj",
+      "model.vision_tower.encoder.layer.19.attention.v_proj",
+      "model.vision_tower.encoder.layer.19.attention.projection_layer",
+      "model.vision_tower.encoder.layer.19.mlp.fc1",
+      "model.vision_tower.encoder.layer.19.mlp.fc2",
+      "model.vision_tower.encoder.layer.20.attention.q_proj",
+      "model.vision_tower.encoder.layer.20.attention.k_proj",
+      "model.vision_tower.encoder.layer.20.attention.v_proj",
+      "model.vision_tower.encoder.layer.20.attention.projection_layer",
+      "model.vision_tower.encoder.layer.20.mlp.fc1",
+      "model.vision_tower.encoder.layer.20.mlp.fc2",
+      "model.vision_tower.encoder.layer.21.attention.q_proj",
+      "model.vision_tower.encoder.layer.21.attention.k_proj",
+      "model.vision_tower.encoder.layer.21.attention.v_proj",
+      "model.vision_tower.encoder.layer.21.attention.projection_layer",
+      "model.vision_tower.encoder.layer.21.mlp.fc1",
+      "model.vision_tower.encoder.layer.21.mlp.fc2",
+      "model.vision_tower.encoder.layer.22.attention.q_proj",
+      "model.vision_tower.encoder.layer.22.attention.k_proj",
+      "model.vision_tower.encoder.layer.22.attention.v_proj",
+      "model.vision_tower.encoder.layer.22.attention.projection_layer",
+      "model.vision_tower.encoder.layer.22.mlp.fc1",
+      "model.vision_tower.encoder.layer.22.mlp.fc2",
+      "model.vision_tower.encoder.layer.23.attention.q_proj",
+      "model.vision_tower.encoder.layer.23.attention.k_proj",
+      "model.vision_tower.encoder.layer.23.attention.v_proj",
+      "model.vision_tower.encoder.layer.23.attention.projection_layer",
+      "model.vision_tower.encoder.layer.23.mlp.fc1",
+      "model.vision_tower.encoder.layer.23.mlp.fc2",
+      "model.multi_modal_projector.linear_1",
+      "model.multi_modal_projector.linear_2",
+      "lm_head"
+    ],
+    "kv_cache_scheme": null,
+    "quant_method": "compressed-tensors",
+    "quantization_status": "compressed",
+    "sparsity_config": {},
+    "transform_config": {},
+    "version": "0.17.1"
+  },
+  "text_config": {
+    "_name_or_path": "/root/codespace/checkpoints/Qwen3-4B",
+    "architectures": [
+      "Qwen3ForCausalLM"
+    ],
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "bos_token_id": 151643,
+    "debug": false,
+    "dtype": "bfloat16",
+    "eos_token_id": 151645,
+    "ep_size": 1,
+    "head_dim": 128,
+    "hidden_act": "silu",
+    "hidden_size": 2560,
+    "initializer_range": 0.02,
+    "intermediate_size": 9728,
+    "layer_types": [
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention"
+    ],
+    "max_position_embeddings": 40960,
+    "max_window_layers": 36,
+    "micro_forward": false,
+    "model_type": "qwen3",
+    "num_attention_heads": 32,
+    "num_hidden_layers": 36,
+    "num_key_value_heads": 8,
+    "pad_token_id": null,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "rope_theta": 1000000,
+      "rope_type": "default"
+    },
+    "skip_checkpoint": false,
+    "sliding_window": null,
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "use_deepep": false,
+    "use_sliding_window": false,
+    "vocab_size": 151936
+  },
+  "tie_word_embeddings": false,
+  "transformers_version": "5.10.1",
+  "vision_config": {
+    "architectures": [
+      "InternVisionModel"
+    ],
+    "attention_bias": true,
+    "attention_dropout": 0.0,
+    "dropout": 0.0,
+    "dtype": "bfloat16",
+    "hidden_act": "gelu",
+    "hidden_dropout_prob": 0.0,
+    "hidden_size": 1024,
+    "image_size": [
+      448,
+      448
+    ],
+    "initializer_factor": 0.1,
+    "initializer_range": 1e-10,
+    "intermediate_size": 4096,
+    "layer_norm_eps": 1e-06,
+    "layer_scale_init_value": 0.1,
+    "model_type": "internvl_vision",
+    "norm_type": "layer_norm",
+    "num_attention_heads": 16,
+    "num_channels": 3,
+    "num_hidden_layers": 24,
+    "patch_size": [
+      14,
+      14
+    ],
+    "projection_dropout": 0.0,
+    "use_absolute_position_embeddings": true,
+    "use_mask_token": false,
+    "use_mean_pooling": true,
+    "use_qk_norm": false
+  },
+  "vision_feature_layer": -1,
+  "vision_feature_select_strategy": "default"
+}

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "transformers_version": "5.10.1"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c42c8a90d5c2c4698e7e043a0d4625ea8df4ef3b8a290f6471064112c5c0a48
+size 4086177728

processor_config.json

@@ -0,0 +1,79 @@
+{
+  "image_processor": {
+    "crop_to_patches": false,
+    "data_format": "channels_first",
+    "default_to_square": true,
+    "do_convert_rgb": true,
+    "do_normalize": true,
+    "do_rescale": true,
+    "do_resize": true,
+    "image_mean": [
+      0.485,
+      0.456,
+      0.406
+    ],
+    "image_processor_type": "GotOcr2ImageProcessor",
+    "image_std": [
+      0.229,
+      0.224,
+      0.225
+    ],
+    "max_patches": 12,
+    "min_patches": 1,
+    "resample": 3,
+    "rescale_factor": 0.00392156862745098,
+    "size": {
+      "height": 448,
+      "width": 448
+    }
+  },
+  "image_seq_length": 256,
+  "processor_class": "InternVLProcessor",
+  "video_processor": {
+    "data_format": "channels_first",
+    "default_to_square": true,
+    "do_convert_rgb": true,
+    "do_normalize": true,
+    "do_rescale": true,
+    "do_resize": true,
+    "do_sample_frames": false,
+    "image_mean": [
+      0.48145466,
+      0.4578275,
+      0.40821073
+    ],
+    "image_std": [
+      0.26862954,
+      0.26130258,
+      0.27577711
+    ],
+    "initial_shift": true,
+    "model_valid_processing_keys": [
+      "do_convert_rgb",
+      "do_resize",
+      "size",
+      "size_divisor",
+      "default_to_square",
+      "resample",
+      "do_rescale",
+      "rescale_factor",
+      "do_normalize",
+      "image_mean",
+      "image_std",
+      "do_pad",
+      "do_center_crop",
+      "crop_size",
+      "data_format",
+      "input_data_format",
+      "device"
+    ],
+    "resample": 3,
+    "rescale_factor": 0.00392156862745098,
+    "return_metadata": false,
+    "size": {
+      "height": 384,
+      "width": 384
+    },
+    "video_processor_type": "InternVLVideoProcessor"
+  }
+}

quantization/recipe.py

@@ -0,0 +1,61 @@
+"""Compression recipe used to create this checkpoint.
+
+Run quantization from the original BF16 checkpoint, not from an already
+quantized checkpoint.
+
+AWQ needs calibration data. This build used 128 samples from
+`lmms-lab/flickr30k`, each rendered at 448x448 (a single InternVL patch) with
+generic English/Korean scene-description prompts. Re-run with in-domain images
+if your target distribution is far from everyday photography.
+
+Targets are regex-scoped to the language decoder, so the vision tower,
+multimodal projector, input embeddings, and lm_head are never matched and stay
+in BF16. `ignore` is therefore empty by design.
+"""
+
+from llmcompressor.modifiers.awq import AWQModifier
+from llmcompressor.modifiers.quantization import QuantizationModifier
+
+ATTENTION = r"re:^model\.language_model\.layers\.\d+\.self_attn\.(q_proj|k_proj|v_proj|o_proj)$"
+MLP = r"re:^model\.language_model\.layers\.\d+\.mlp\.(gate_proj|up_proj|down_proj)$"
+
+RECIPE = [
+    AWQModifier(
+        mappings=[
+            {
+                "smooth_layer": r"re:^model\.language_model\.layers\.\d+\.input_layernorm$",
+                "balance_layers": [
+                    r"re:^model\.language_model\.layers\.\d+\.self_attn\.q_proj$",
+                    r"re:^model\.language_model\.layers\.\d+\.self_attn\.k_proj$",
+                    r"re:^model\.language_model\.layers\.\d+\.self_attn\.v_proj$",
+                ],
+            },
+            {
+                "smooth_layer": r"re:^model\.language_model\.layers\.\d+\.self_attn\.v_proj$",
+                "balance_layers": [
+                    r"re:^model\.language_model\.layers\.\d+\.self_attn\.o_proj$",
+                ],
+            },
+            {
+                "smooth_layer": r"re:^model\.language_model\.layers\.\d+\.post_attention_layernorm$",
+                "balance_layers": [
+                    r"re:^model\.language_model\.layers\.\d+\.mlp\.gate_proj$",
+                    r"re:^model\.language_model\.layers\.\d+\.mlp\.up_proj$",
+                ],
+            },
+            {
+                "smooth_layer": r"re:^model\.language_model\.layers\.\d+\.mlp\.up_proj$",
+                "balance_layers": [
+                    r"re:^model\.language_model\.layers\.\d+\.mlp\.down_proj$",
+                ],
+            },
+        ],
+        duo_scaling=True,
+        n_grid=20,
+    ),
+    QuantizationModifier(
+        targets=[ATTENTION, MLP],
+        ignore=[],
+        scheme="W4A16_ASYM",
+    ),
+]

recipe.yaml

@@ -0,0 +1,25 @@
+default_stage:
+  default_modifiers:
+    AWQModifier:
+      mappings:
+      - smooth_layer: re:^model\.language_model\.layers\.\d+\.input_layernorm$
+        balance_layers: ['re:^model\.language_model\.layers\.\d+\.self_attn\.q_proj$', 're:^model\.language_model\.layers\.\d+\.self_attn\.k_proj$',
+          're:^model\.language_model\.layers\.\d+\.self_attn\.v_proj$']
+        activation_hook_target: null
+      - smooth_layer: re:^model\.language_model\.layers\.\d+\.self_attn\.v_proj$
+        balance_layers: ['re:^model\.language_model\.layers\.\d+\.self_attn\.o_proj$']
+        activation_hook_target: null
+      - smooth_layer: re:^model\.language_model\.layers\.\d+\.post_attention_layernorm$
+        balance_layers: ['re:^model\.language_model\.layers\.\d+\.mlp\.gate_proj$', 're:^model\.language_model\.layers\.\d+\.mlp\.up_proj$']
+        activation_hook_target: null
+      - smooth_layer: re:^model\.language_model\.layers\.\d+\.mlp\.up_proj$
+        balance_layers: ['re:^model\.language_model\.layers\.\d+\.mlp\.down_proj$']
+        activation_hook_target: null
+      duo_scaling: true
+      n_grid: 20
+    QuantizationModifier:
+      targets: ['re:^model\.language_model\.layers\.\d+\.self_attn\.(q_proj|k_proj|v_proj|o_proj)$',
+        're:^model\.language_model\.layers\.\d+\.mlp\.(gate_proj|up_proj|down_proj)$']
+      ignore: []
+      scheme: W4A16_ASYM
+      bypass_divisibility_checks: false

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b9d18660f656ae5a87df2d5d6ed990e80f292d3473c1a35cae8259a5d28cd67
+size 11424484

tokenizer_config.json

@@ -0,0 +1,26 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "context_image_token": "<IMG_CONTEXT>",
+  "end_image_token": "</img>",
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "is_local": true,
+  "local_files_only": true,
+  "model_max_length": 40960,
+  "model_specific_special_tokens": {
+    "context_image_token": "<IMG_CONTEXT>",
+    "end_image_token": "</img>",
+    "start_image_token": "<img>",
+    "video_token": "<video>"
+  },
+  "pad_token": "<|endoftext|>",
+  "processor_class": "InternVLProcessor",
+  "split_special_tokens": false,
+  "start_image_token": "<img>",
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<video>"
+}