Image-Text-to-Text
Safetensors
Trellis
vllm
qwen3_5
exl3
exllamav3
mixed-precision
quantized
qwen3.8
vision-language
gilded-gnosis
conversational
4-bit precision
Instructions to use malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Trellis
How to use malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated with Trellis:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Add pinned QFS size–KL plot, highlighted deep link and downloadable provenance
Browse files- DOCS-SHA256SUMS +5 -1
- README.md +12 -0
- assets/qfs-size-kl-v5.csv +7 -0
- assets/qfs-size-kl-v5.json +308 -0
- assets/qfs-size-kl-v5.png +0 -0
- assets/qfs-size-kl-v5.svg +11 -0
DOCS-SHA256SUMS
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
0b379462fd9cb8fbbbfbabd3d4d9c03f89aeb357dc846c8e6d74ffad596229b2 .gitattributes
|
| 2 |
-
|
| 3 |
d71eb992953ffd25e07d268a268dbf4953cae40eb14e35ef0d3d7e12f32709c2 assets/context-frontier-dark.png
|
| 4 |
df4d8cfdb5ded9e330edbc16223d83ca524f3b67e7bcea910c5a7b4530f11b1b assets/context-frontier-dark.svg
|
| 5 |
bafed8e1de217f9311dbf604ab69950adb00e8f3d23a9c2a3ea12184cb47ac8a assets/context-frontier-light.png
|
|
@@ -20,4 +20,8 @@ b4c9052ee362a31890338fc4427af334d7eb6dfbaa1fd7572bf85832fcccb81e assets/kld-fam
|
|
| 20 |
1c3cd2bebbe9addad05787136ca307da8f894093c2a58105b460238c65781bb5 assets/kld-family-comparison-dark.svg
|
| 21 |
5f1f9b92cabb1086bfabcff24b7d215bdd95430ff829822759bf04e89e02884f assets/kld-family-comparison-light.png
|
| 22 |
3e9d0bf26c4b83dfce205d4e1486ffef6441678a24d7d13b0de2917c64db3ffb assets/kld-family-comparison-light.svg
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
0602ce8b2ba9e95b84adc5933ab1138faec605a6c0b12af935bbcd91d94d6aff release-evidence.json
|
|
|
|
| 1 |
0b379462fd9cb8fbbbfbabd3d4d9c03f89aeb357dc846c8e6d74ffad596229b2 .gitattributes
|
| 2 |
+
a9d6ac3c430f9f70708a7b5f21edd7fa1af991e6b0ebeb213cfaf04b341213d4 README.md
|
| 3 |
d71eb992953ffd25e07d268a268dbf4953cae40eb14e35ef0d3d7e12f32709c2 assets/context-frontier-dark.png
|
| 4 |
df4d8cfdb5ded9e330edbc16223d83ca524f3b67e7bcea910c5a7b4530f11b1b assets/context-frontier-dark.svg
|
| 5 |
bafed8e1de217f9311dbf604ab69950adb00e8f3d23a9c2a3ea12184cb47ac8a assets/context-frontier-light.png
|
|
|
|
| 20 |
1c3cd2bebbe9addad05787136ca307da8f894093c2a58105b460238c65781bb5 assets/kld-family-comparison-dark.svg
|
| 21 |
5f1f9b92cabb1086bfabcff24b7d215bdd95430ff829822759bf04e89e02884f assets/kld-family-comparison-light.png
|
| 22 |
3e9d0bf26c4b83dfce205d4e1486ffef6441678a24d7d13b0de2917c64db3ffb assets/kld-family-comparison-light.svg
|
| 23 |
+
e14bd8b12131f6a8af7ea7c736d22265357a9d78df5ad943b0eb51217f21fd18 assets/qfs-size-kl-v5.csv
|
| 24 |
+
9c5f8eafea6777e7cfc8ba89d863a55e29774883f9d07e1e5dbff62835f8df73 assets/qfs-size-kl-v5.json
|
| 25 |
+
827dfb12cfc9f719e9b5ce8f7f6af933b61821abdfeee6bcc28444aaae9581e1 assets/qfs-size-kl-v5.png
|
| 26 |
+
2aa648fc9d9d42b2cbd28fa4f288d1bc6a27bfe8297fd9c41eb471341b77ac0b assets/qfs-size-kl-v5.svg
|
| 27 |
0602ce8b2ba9e95b84adc5933ab1138faec605a6c0b12af935bbcd91d94d6aff release-evidence.json
|
README.md
CHANGED
|
@@ -61,6 +61,18 @@ read-only images, ephemeral containers, many nodes. **Choose the sibling** if yo
|
|
| 61 |
fidelity for KV room, and on a 32 GB card that difference matters (see
|
| 62 |
[context](#context-capacity)).
|
| 63 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
## Which of the four builds
|
| 65 |
|
| 66 |
Same architecture and tokenizer. The headline KLD column is the **v5 held-out suite,
|
|
|
|
| 61 |
fidelity for KV room, and on a 32 GB card that difference matters (see
|
| 62 |
[context](#context-capacity)).
|
| 63 |
|
| 64 |
+
<!-- QFS-SIZE-KL-BEGIN -->
|
| 65 |
+
## QFS size–fidelity evidence
|
| 66 |
+
|
| 67 |
+
[](https://malaiwah-qfs-explorer.hf.space/?tab=plots&measurement=measurement--qwen38.k5k6-hydrated.suite-v5-10m®istry_revision=ae6a3fa203aede6e8aea20fba3e10ac3b74930ce&scale=auto)
|
| 68 |
+
|
| 69 |
+
The v5 10M-position **shared-head, body-only** registry group; this model is highlighted without hiding its peers. The x-axis uses recorded serialized bytes with the basis disclosed, **not VRAM or a live download-size estimate**. This group is **inspection-only**, not a certified ranking: the registry retains its scope/provenance limitations and draws no ranking line. Unknown-size peers remain listed in the downloadable data, not assigned invented sizes.
|
| 70 |
+
|
| 71 |
+
[Interactive highlighted plot](https://malaiwah-qfs-explorer.hf.space/?tab=plots&measurement=measurement--qwen38.k5k6-hydrated.suite-v5-10m®istry_revision=ae6a3fa203aede6e8aea20fba3e10ac3b74930ce&scale=auto) · [PNG](assets/qfs-size-kl-v5.png) · [SVG](assets/qfs-size-kl-v5.svg) · [CSV and exclusions](assets/qfs-size-kl-v5.csv) · [Full provenance JSON](assets/qfs-size-kl-v5.json) · [Live cached image](https://malaiwah-qfs-explorer.hf.space/plots/live.png?measurement=measurement--qwen38.k5k6-hydrated.suite-v5-10m&scale=auto)
|
| 72 |
+
|
| 73 |
+
Registry snapshot: `ae6a3fa203aede6e8aea20fba3e10ac3b74930ce`. The static image and data above are stored with this card; the live image is explicitly mutable. No model weights or fidelity values were changed by this card update.
|
| 74 |
+
<!-- QFS-SIZE-KL-END -->
|
| 75 |
+
|
| 76 |
## Which of the four builds
|
| 77 |
|
| 78 |
Same architecture and tokenizer. The headline KLD column is the **v5 held-out suite,
|
assets/qfs-size-kl-v5.csv
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
measurement_id,artifact_name,classification,is_control,is_highlight,plotted,exclusion_reason,size_bytes,size_gib,size_basis,kl_nats,kl_display,y_unit,scale,scale_note,artifact_revision,source_url,dataset_viewer_url,group_id,group_status,ranking_allowed,panel_id,panel_name,reference_id,reference_name,registry_revision,registry_origin
|
| 2 |
+
measurement--qwen38.k4.suite-v5-10m,malaiwah Qwen3.8-27B K4,advisory,False,False,True,,28345369355,26.3987,repo_all_files,0.0106039,10.6039,millinats,linear,,,https://huggingface.co/malaiwah/Qwen3.8-27B-K4,https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.k4.suite-v5-10m,"[""cmp--c8c4df32774bdb63"",""""]",false,False,panel--qwen38.malaiwah.suite-v5-10m,"malaiwah Qwen3.8-27B distribution-fidelity suite v5 -- 5,120 contexts",reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m,malaiwah Qwen3.8-27B BF16 shared-head capture over qwen38.malaiwah.suite-v5-10m,ae6a3fa203aede6e8aea20fba3e10ac3b74930ce,fetched from the public HF dataset malaiwah/quant-fidelity-registry
|
| 3 |
+
measurement--qwen38.k5k6-context.suite-v5-10m,malaiwah Qwen3.8-27B EXL3 K5K6 context,advisory,False,False,True,,20696053306,19.2747,repo_all_files,0.00350936,3.50936,millinats,linear,,,https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-context,https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.k5k6-context.suite-v5-10m,"[""cmp--c8c4df32774bdb63"",""""]",false,False,panel--qwen38.malaiwah.suite-v5-10m,"malaiwah Qwen3.8-27B distribution-fidelity suite v5 -- 5,120 contexts",reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m,malaiwah Qwen3.8-27B BF16 shared-head capture over qwen38.malaiwah.suite-v5-10m,ae6a3fa203aede6e8aea20fba3e10ac3b74930ce,fetched from the public HF dataset malaiwah/quant-fidelity-registry
|
| 4 |
+
measurement--qwen38.k5k6-hydrated.suite-v5-10m,malaiwah Qwen3.8-27B EXL3 K5K6 hydrated,advisory,False,True,True,,21610933884,20.1268,repo_all_files,0.00275963,2.75963,millinats,linear,,,https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated,https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.k5k6-hydrated.suite-v5-10m,"[""cmp--c8c4df32774bdb63"",""""]",false,False,panel--qwen38.malaiwah.suite-v5-10m,"malaiwah Qwen3.8-27B distribution-fidelity suite v5 -- 5,120 contexts",reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m,malaiwah Qwen3.8-27B BF16 shared-head capture over qwen38.malaiwah.suite-v5-10m,ae6a3fa203aede6e8aea20fba3e10ac3b74930ce,fetched from the public HF dataset malaiwah/quant-fidelity-registry
|
| 5 |
+
measurement--qwen38.k5k6.suite-v5-10m,malaiwah Qwen3.8-27B EXL3 K5K6,advisory,False,False,True,,30597231933,28.4959,repo_all_files,0.00320988,3.20988,millinats,linear,,,https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6,https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.k5k6.suite-v5-10m,"[""cmp--c8c4df32774bdb63"",""""]",false,False,panel--qwen38.malaiwah.suite-v5-10m,"malaiwah Qwen3.8-27B distribution-fidelity suite v5 -- 5,120 contexts",reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m,malaiwah Qwen3.8-27B BF16 shared-head capture over qwen38.malaiwah.suite-v5-10m,ae6a3fa203aede6e8aea20fba3e10ac3b74930ce,fetched from the public HF dataset malaiwah/quant-fidelity-registry
|
| 6 |
+
measurement--qwen38.official-fp8.suite-v5-10m,Qwen3.8-27B FP8 (official),advisory,False,False,True,,30890049597,28.7686,repo_all_files,0.00529394,5.29394,millinats,linear,,,https://huggingface.co/Qwen/Qwen3.8-27B-FP8,https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.official-fp8.suite-v5-10m,"[""cmp--c8c4df32774bdb63"",""""]",false,False,panel--qwen38.malaiwah.suite-v5-10m,"malaiwah Qwen3.8-27B distribution-fidelity suite v5 -- 5,120 contexts",reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m,malaiwah Qwen3.8-27B BF16 shared-head capture over qwen38.malaiwah.suite-v5-10m,ae6a3fa203aede6e8aea20fba3e10ac3b74930ce,fetched from the public HF dataset malaiwah/quant-fidelity-registry
|
| 7 |
+
measurement--qwen38.unsloth-nvfp4.suite-v5-10m,unsloth Qwen3.8-27B NVFP4,advisory,False,False,False,serialized size not recorded,,,unknown,,,millinats,linear,,9c73e2daee1d0fd494ffbd1d8753f2174a953796,https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4,https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.unsloth-nvfp4.suite-v5-10m,"[""cmp--c8c4df32774bdb63"",""""]",false,False,panel--qwen38.malaiwah.suite-v5-10m,"malaiwah Qwen3.8-27B distribution-fidelity suite v5 -- 5,120 contexts",reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m,malaiwah Qwen3.8-27B BF16 shared-head capture over qwen38.malaiwah.suite-v5-10m,ae6a3fa203aede6e8aea20fba3e10ac3b74930ce,fetched from the public HF dataset malaiwah/quant-fidelity-registry
|
assets/qfs-size-kl-v5.json
ADDED
|
@@ -0,0 +1,308 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema": "qfs-explorer-plots/v1",
|
| 3 |
+
"kind": "scatter",
|
| 4 |
+
"title": "Qwen3.8-27B \u2014 size vs KL",
|
| 5 |
+
"subtitle": "Panel: malaiwah Qwen3.8-27B distribution-fidelity suite v5 -- 5,120 contexts \u00b7 Reference: malaiwah Qwen3.8-27B BF16 shared-head capture over qwen38.malaiwah.suite-v5-10m",
|
| 6 |
+
"caution": "Lower KL is closer to the reference, only within this comparability group. x is recorded serialized size (GiB), not VRAM. Size basis is retained per point; unknown sizes and missing/non-finite KL are excluded with reasons. Group comparability is false; this plot is inspection-only and not a ranking.",
|
| 7 |
+
"scale_requested": "auto",
|
| 8 |
+
"scale": "linear",
|
| 9 |
+
"scale_note": "",
|
| 10 |
+
"y_unit": "millinats",
|
| 11 |
+
"y_unit_factor": 1000.0,
|
| 12 |
+
"registry_revision": "ae6a3fa203aede6e8aea20fba3e10ac3b74930ce",
|
| 13 |
+
"registry_snapshot": "ae6a3fa203ae",
|
| 14 |
+
"registry_origin": "fetched from the public HF dataset malaiwah/quant-fidelity-registry",
|
| 15 |
+
"registry_notes": [],
|
| 16 |
+
"highlight": {
|
| 17 |
+
"measurement_id": "measurement--qwen38.k5k6-hydrated.suite-v5-10m",
|
| 18 |
+
"name": "malaiwah Qwen3.8-27B EXL3 K5K6 hydrated",
|
| 19 |
+
"excluded": false,
|
| 20 |
+
"exclusion_reason": null
|
| 21 |
+
},
|
| 22 |
+
"group": {
|
| 23 |
+
"id": "[\"cmp--c8c4df32774bdb63\",\"\"]",
|
| 24 |
+
"title": "Qwen3.8-27B \u2014 size vs KL",
|
| 25 |
+
"status": "false",
|
| 26 |
+
"lane": null,
|
| 27 |
+
"key": "cmp--c8c4df32774bdb63",
|
| 28 |
+
"ranking_allowed": false,
|
| 29 |
+
"ordering": "measurement identity; not a ranking",
|
| 30 |
+
"predicate": {
|
| 31 |
+
"predicate_version": "comparability-predicate/v2",
|
| 32 |
+
"comparable": "false",
|
| 33 |
+
"reasons": [
|
| 34 |
+
"harness: unrecorded for 6 members; homogeneity cannot be certified",
|
| 35 |
+
"stack: unrecorded for 6 members; homogeneity cannot be certified",
|
| 36 |
+
"replay_backend: unrecorded for 6 members; homogeneity cannot be certified",
|
| 37 |
+
"replay_env: unrecorded for 6 members; homogeneity cannot be certified",
|
| 38 |
+
"scope: members span {attn.o+attn.qkv+lm_head+mlp.down+mlp.gate+mlp.up+mtp|head=quantized|kv=bf16, attn.o+attn.qkv+mlp.down+mlp.gate+mlp.up|head=native|kv=bf16, lm_head+mlp.down+mlp.gate+mlp.up|head=quantized|kv=bf16, unrecorded} -- these are different interventions; equal nominal bpw does not make a routed-experts-only quant and a full-forward quant the same thing (docs/GGUF-MEASUREMENT.md)"
|
| 39 |
+
],
|
| 40 |
+
"secondary": {
|
| 41 |
+
"key": {
|
| 42 |
+
"status": "pass",
|
| 43 |
+
"values": [
|
| 44 |
+
"cmp--c8c4df32774bdb63"
|
| 45 |
+
]
|
| 46 |
+
},
|
| 47 |
+
"harness": {
|
| 48 |
+
"status": "unknown",
|
| 49 |
+
"values": [
|
| 50 |
+
"unrecorded"
|
| 51 |
+
]
|
| 52 |
+
},
|
| 53 |
+
"stack": {
|
| 54 |
+
"status": "unknown",
|
| 55 |
+
"values": [
|
| 56 |
+
"unrecorded"
|
| 57 |
+
]
|
| 58 |
+
},
|
| 59 |
+
"replay_backend": {
|
| 60 |
+
"status": "unknown",
|
| 61 |
+
"values": [
|
| 62 |
+
"unrecorded"
|
| 63 |
+
]
|
| 64 |
+
},
|
| 65 |
+
"replay_env": {
|
| 66 |
+
"status": "unknown",
|
| 67 |
+
"values": [
|
| 68 |
+
"unrecorded"
|
| 69 |
+
]
|
| 70 |
+
},
|
| 71 |
+
"lane": {
|
| 72 |
+
"status": "pass",
|
| 73 |
+
"values": [
|
| 74 |
+
"sealed-ep8"
|
| 75 |
+
]
|
| 76 |
+
},
|
| 77 |
+
"pipeline": {
|
| 78 |
+
"status": "pass",
|
| 79 |
+
"values": [
|
| 80 |
+
"pipeline--malaiwah.qwen38-kld-ladder"
|
| 81 |
+
]
|
| 82 |
+
},
|
| 83 |
+
"scope": {
|
| 84 |
+
"status": "fail",
|
| 85 |
+
"values": [
|
| 86 |
+
"attn.o+attn.qkv+lm_head+mlp.down+mlp.gate+mlp.up+mtp|head=quantized|kv=bf16",
|
| 87 |
+
"attn.o+attn.qkv+mlp.down+mlp.gate+mlp.up|head=native|kv=bf16",
|
| 88 |
+
"lm_head+mlp.down+mlp.gate+mlp.up|head=quantized|kv=bf16",
|
| 89 |
+
"unrecorded"
|
| 90 |
+
]
|
| 91 |
+
},
|
| 92 |
+
"hardware": {
|
| 93 |
+
"status": "pass",
|
| 94 |
+
"values": [
|
| 95 |
+
"cuda"
|
| 96 |
+
]
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"live_member_count": 6
|
| 100 |
+
},
|
| 101 |
+
"predicate_scope": "all original members of the full comparability key, before lane or model filtering",
|
| 102 |
+
"reasons": [
|
| 103 |
+
"harness: unrecorded for 6 members; homogeneity cannot be certified",
|
| 104 |
+
"stack: unrecorded for 6 members; homogeneity cannot be certified",
|
| 105 |
+
"replay_backend: unrecorded for 6 members; homogeneity cannot be certified",
|
| 106 |
+
"replay_env: unrecorded for 6 members; homogeneity cannot be certified",
|
| 107 |
+
"scope: members span {attn.o+attn.qkv+lm_head+mlp.down+mlp.gate+mlp.up+mtp|head=quantized|kv=bf16, attn.o+attn.qkv+mlp.down+mlp.gate+mlp.up|head=native|kv=bf16, lm_head+mlp.down+mlp.gate+mlp.up|head=quantized|kv=bf16, unrecorded} -- these are different interventions; equal nominal bpw does not make a routed-experts-only quant and a full-forward quant the same thing (docs/GGUF-MEASUREMENT.md)"
|
| 108 |
+
],
|
| 109 |
+
"caveats": [],
|
| 110 |
+
"rule": "Different keys or lanes are never ranked together. Equal keys alone do not certify like-for-like comparison. Comparability is not a claim of general model quality.",
|
| 111 |
+
"snapshot_notes": [],
|
| 112 |
+
"key_inputs": {
|
| 113 |
+
"panel_id": "panel--qwen38.malaiwah.suite-v5-10m",
|
| 114 |
+
"reference_id": "reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m",
|
| 115 |
+
"metric_name": "mean_tokenwise_kld",
|
| 116 |
+
"direction": "reference_to_candidate",
|
| 117 |
+
"accumulation_dtype": "float32_reduce_legacy",
|
| 118 |
+
"stack_relation": "same_stack",
|
| 119 |
+
"head_policy": "shared_reference_head"
|
| 120 |
+
},
|
| 121 |
+
"full_key_member_count": 6,
|
| 122 |
+
"displayed_member_count": 6,
|
| 123 |
+
"control_measurement_ids": [],
|
| 124 |
+
"member_count_total": 6,
|
| 125 |
+
"member_count_plotted": 5,
|
| 126 |
+
"member_count_excluded": 1
|
| 127 |
+
},
|
| 128 |
+
"panel": {
|
| 129 |
+
"id": "panel--qwen38.malaiwah.suite-v5-10m",
|
| 130 |
+
"name": "malaiwah Qwen3.8-27B distribution-fidelity suite v5 -- 5,120 contexts"
|
| 131 |
+
},
|
| 132 |
+
"reference": {
|
| 133 |
+
"id": "reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m",
|
| 134 |
+
"name": "malaiwah Qwen3.8-27B BF16 shared-head capture over qwen38.malaiwah.suite-v5-10m"
|
| 135 |
+
},
|
| 136 |
+
"model": {
|
| 137 |
+
"name": "Qwen3.8-27B"
|
| 138 |
+
},
|
| 139 |
+
"axes": {
|
| 140 |
+
"x": {
|
| 141 |
+
"min": 0.0,
|
| 142 |
+
"max": 28.768600520677865,
|
| 143 |
+
"label": "Serialized size (GiB; basis disclosed)",
|
| 144 |
+
"scale": "linear",
|
| 145 |
+
"ticks": [
|
| 146 |
+
0,
|
| 147 |
+
10,
|
| 148 |
+
20
|
| 149 |
+
],
|
| 150 |
+
"tick_labels": [
|
| 151 |
+
"0",
|
| 152 |
+
"10",
|
| 153 |
+
"20"
|
| 154 |
+
]
|
| 155 |
+
},
|
| 156 |
+
"y": {
|
| 157 |
+
"min": 0.0,
|
| 158 |
+
"max": 10.603886897374757,
|
| 159 |
+
"label": "Mean KL (millinats)",
|
| 160 |
+
"scale": "linear",
|
| 161 |
+
"linthresh": null,
|
| 162 |
+
"ticks": [
|
| 163 |
+
0,
|
| 164 |
+
5,
|
| 165 |
+
10
|
| 166 |
+
],
|
| 167 |
+
"tick_labels": [
|
| 168 |
+
"0",
|
| 169 |
+
"5",
|
| 170 |
+
"10"
|
| 171 |
+
]
|
| 172 |
+
}
|
| 173 |
+
},
|
| 174 |
+
"points": [
|
| 175 |
+
{
|
| 176 |
+
"measurement_id": "measurement--qwen38.k4.suite-v5-10m",
|
| 177 |
+
"artifact": "malaiwah Qwen3.8-27B K4",
|
| 178 |
+
"name": "malaiwah Qwen3.8-27B K4",
|
| 179 |
+
"label": "K4",
|
| 180 |
+
"classification": "advisory",
|
| 181 |
+
"is_control": false,
|
| 182 |
+
"is_highlight": false,
|
| 183 |
+
"revision": null,
|
| 184 |
+
"size_bytes": 28345369355,
|
| 185 |
+
"size_basis": "repo_all_files",
|
| 186 |
+
"size_gib": 26.398682366125286,
|
| 187 |
+
"kl_nats": 0.010603886897374758,
|
| 188 |
+
"source_url": "https://huggingface.co/malaiwah/Qwen3.8-27B-K4",
|
| 189 |
+
"dataset_viewer_url": "https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.k4.suite-v5-10m",
|
| 190 |
+
"plotted": true,
|
| 191 |
+
"exclusion_reason": null,
|
| 192 |
+
"kl_display": 10.603886897374757
|
| 193 |
+
},
|
| 194 |
+
{
|
| 195 |
+
"measurement_id": "measurement--qwen38.k5k6-context.suite-v5-10m",
|
| 196 |
+
"artifact": "malaiwah Qwen3.8-27B EXL3 K5K6 context",
|
| 197 |
+
"name": "malaiwah Qwen3.8-27B EXL3 K5K6 context",
|
| 198 |
+
"label": "EXL3 K5K6 context",
|
| 199 |
+
"classification": "advisory",
|
| 200 |
+
"is_control": false,
|
| 201 |
+
"is_highlight": false,
|
| 202 |
+
"revision": null,
|
| 203 |
+
"size_bytes": 20696053306,
|
| 204 |
+
"size_basis": "repo_all_files",
|
| 205 |
+
"size_gib": 19.274701649323106,
|
| 206 |
+
"kl_nats": 0.0035093587497007383,
|
| 207 |
+
"source_url": "https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-context",
|
| 208 |
+
"dataset_viewer_url": "https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.k5k6-context.suite-v5-10m",
|
| 209 |
+
"plotted": true,
|
| 210 |
+
"exclusion_reason": null,
|
| 211 |
+
"kl_display": 3.509358749700738
|
| 212 |
+
},
|
| 213 |
+
{
|
| 214 |
+
"measurement_id": "measurement--qwen38.k5k6-hydrated.suite-v5-10m",
|
| 215 |
+
"artifact": "malaiwah Qwen3.8-27B EXL3 K5K6 hydrated",
|
| 216 |
+
"name": "malaiwah Qwen3.8-27B EXL3 K5K6 hydrated",
|
| 217 |
+
"label": "EXL3 K5K6 hydrated",
|
| 218 |
+
"classification": "advisory",
|
| 219 |
+
"is_control": false,
|
| 220 |
+
"is_highlight": true,
|
| 221 |
+
"revision": null,
|
| 222 |
+
"size_bytes": 21610933884,
|
| 223 |
+
"size_basis": "repo_all_files",
|
| 224 |
+
"size_gib": 20.126750584691763,
|
| 225 |
+
"kl_nats": 0.0027596250151498245,
|
| 226 |
+
"source_url": "https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated",
|
| 227 |
+
"dataset_viewer_url": "https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.k5k6-hydrated.suite-v5-10m",
|
| 228 |
+
"plotted": true,
|
| 229 |
+
"exclusion_reason": null,
|
| 230 |
+
"kl_display": 2.7596250151498243
|
| 231 |
+
},
|
| 232 |
+
{
|
| 233 |
+
"measurement_id": "measurement--qwen38.k5k6.suite-v5-10m",
|
| 234 |
+
"artifact": "malaiwah Qwen3.8-27B EXL3 K5K6",
|
| 235 |
+
"name": "malaiwah Qwen3.8-27B EXL3 K5K6",
|
| 236 |
+
"label": "EXL3 K5K6",
|
| 237 |
+
"classification": "advisory",
|
| 238 |
+
"is_control": false,
|
| 239 |
+
"is_highlight": false,
|
| 240 |
+
"revision": null,
|
| 241 |
+
"size_bytes": 30597231933,
|
| 242 |
+
"size_basis": "repo_all_files",
|
| 243 |
+
"size_gib": 28.495892819948494,
|
| 244 |
+
"kl_nats": 0.0032098839572023935,
|
| 245 |
+
"source_url": "https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6",
|
| 246 |
+
"dataset_viewer_url": "https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.k5k6.suite-v5-10m",
|
| 247 |
+
"plotted": true,
|
| 248 |
+
"exclusion_reason": null,
|
| 249 |
+
"kl_display": 3.2098839572023934
|
| 250 |
+
},
|
| 251 |
+
{
|
| 252 |
+
"measurement_id": "measurement--qwen38.official-fp8.suite-v5-10m",
|
| 253 |
+
"artifact": "Qwen3.8-27B FP8 (official)",
|
| 254 |
+
"name": "Qwen3.8-27B FP8 (official)",
|
| 255 |
+
"label": "FP8 (official)",
|
| 256 |
+
"classification": "advisory",
|
| 257 |
+
"is_control": false,
|
| 258 |
+
"is_highlight": false,
|
| 259 |
+
"revision": null,
|
| 260 |
+
"size_bytes": 30890049597,
|
| 261 |
+
"size_basis": "repo_all_files",
|
| 262 |
+
"size_gib": 28.768600520677865,
|
| 263 |
+
"kl_nats": 0.005293941260388515,
|
| 264 |
+
"source_url": "https://huggingface.co/Qwen/Qwen3.8-27B-FP8",
|
| 265 |
+
"dataset_viewer_url": "https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.official-fp8.suite-v5-10m",
|
| 266 |
+
"plotted": true,
|
| 267 |
+
"exclusion_reason": null,
|
| 268 |
+
"kl_display": 5.293941260388515
|
| 269 |
+
}
|
| 270 |
+
],
|
| 271 |
+
"exclusions": [
|
| 272 |
+
{
|
| 273 |
+
"measurement_id": "measurement--qwen38.unsloth-nvfp4.suite-v5-10m",
|
| 274 |
+
"artifact": "unsloth Qwen3.8-27B NVFP4",
|
| 275 |
+
"name": "unsloth Qwen3.8-27B NVFP4",
|
| 276 |
+
"label": "NVFP4",
|
| 277 |
+
"classification": "advisory",
|
| 278 |
+
"is_control": false,
|
| 279 |
+
"is_highlight": false,
|
| 280 |
+
"revision": "9c73e2daee1d0fd494ffbd1d8753f2174a953796",
|
| 281 |
+
"size_bytes": null,
|
| 282 |
+
"size_basis": "unknown",
|
| 283 |
+
"size_gib": null,
|
| 284 |
+
"kl_nats": 0.031058607459514095,
|
| 285 |
+
"source_url": "https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4",
|
| 286 |
+
"dataset_viewer_url": "https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train?q=measurement--qwen38.unsloth-nvfp4.suite-v5-10m",
|
| 287 |
+
"plotted": false,
|
| 288 |
+
"kl_display": null,
|
| 289 |
+
"exclusion_reason": "serialized size not recorded"
|
| 290 |
+
}
|
| 291 |
+
],
|
| 292 |
+
"pareto": {
|
| 293 |
+
"shown": false,
|
| 294 |
+
"order": [],
|
| 295 |
+
"note": ""
|
| 296 |
+
},
|
| 297 |
+
"links": {
|
| 298 |
+
"registry_dataset": "https://huggingface.co/datasets/malaiwah/quant-fidelity-registry",
|
| 299 |
+
"immutable_registry_records": "https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/resolve/ae6a3fa203aede6e8aea20fba3e10ac3b74930ce/data/measurements.jsonl",
|
| 300 |
+
"dataset_viewer_measurements": "https://huggingface.co/datasets/malaiwah/quant-fidelity-registry/viewer/measurements/train",
|
| 301 |
+
"viewer_note": "The dataset viewer searches live data; only the Explorer and raw records above pin this registry snapshot."
|
| 302 |
+
},
|
| 303 |
+
"render": {
|
| 304 |
+
"width": 840,
|
| 305 |
+
"height": 580,
|
| 306 |
+
"max_points": 48
|
| 307 |
+
}
|
| 308 |
+
}
|
assets/qfs-size-kl-v5.png
ADDED
|
assets/qfs-size-kl-v5.svg
ADDED
|
|