Image-Text-to-Text
Safetensors
Merlin
English
vllm
qwen4_exp
qwen3.8
flash-next
w4a16
autoround
mtp
speculative-decoding
fp8-kv
long-context
vision
rtx-3090
ampere
sm86
conversational
compressed-tensors
Instructions to use halt95/Qwen3.8-Flash-Next-W4A16-Merlin with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Merlin
How to use halt95/Qwen3.8-Flash-Next-W4A16-Merlin with Merlin:
# 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
calibrated fp8 KV scale sidecar (2026-09-08)
Browse files- qsa_kv_scales_262k.json +50 -0
qsa_kv_scales_262k.json
ADDED
|
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"language_model.model.layers.11.self_attn.attn": {
|
| 3 |
+
"k_scale": 0.020717075892857144,
|
| 4 |
+
"v_scale": 0.044196428571428574
|
| 5 |
+
},
|
| 6 |
+
"language_model.model.layers.15.self_attn.attn": {
|
| 7 |
+
"k_scale": 0.021637834821428576,
|
| 8 |
+
"v_scale": 0.03069196428571429
|
| 9 |
+
},
|
| 10 |
+
"language_model.model.layers.19.self_attn.attn": {
|
| 11 |
+
"k_scale": 0.024860491071428573,
|
| 12 |
+
"v_scale": 0.06690848214285715
|
| 13 |
+
},
|
| 14 |
+
"language_model.model.layers.23.self_attn.attn": {
|
| 15 |
+
"k_scale": 0.03422154017857143,
|
| 16 |
+
"v_scale": 0.02562779017857143
|
| 17 |
+
},
|
| 18 |
+
"language_model.model.layers.27.self_attn.attn": {
|
| 19 |
+
"k_scale": 0.038825334821428574,
|
| 20 |
+
"v_scale": 0.07857142857142858
|
| 21 |
+
},
|
| 22 |
+
"language_model.model.layers.3.self_attn.attn": {
|
| 23 |
+
"k_scale": 0.018031529017857144,
|
| 24 |
+
"v_scale": 0.012353515625000002
|
| 25 |
+
},
|
| 26 |
+
"language_model.model.layers.31.self_attn.attn": {
|
| 27 |
+
"k_scale": 0.026548549107142858,
|
| 28 |
+
"v_scale": 0.04388950892857143
|
| 29 |
+
},
|
| 30 |
+
"language_model.model.layers.35.self_attn.attn": {
|
| 31 |
+
"k_scale": 0.02639508928571429,
|
| 32 |
+
"v_scale": 0.07765066964285715
|
| 33 |
+
},
|
| 34 |
+
"language_model.model.layers.39.self_attn.attn": {
|
| 35 |
+
"k_scale": 0.037444196428571434,
|
| 36 |
+
"v_scale": 0.03989955357142857
|
| 37 |
+
},
|
| 38 |
+
"language_model.model.layers.43.self_attn.attn": {
|
| 39 |
+
"k_scale": 0.02255859375,
|
| 40 |
+
"v_scale": 0.017647879464285716
|
| 41 |
+
},
|
| 42 |
+
"language_model.model.layers.47.self_attn.attn": {
|
| 43 |
+
"k_scale": 0.03897879464285715,
|
| 44 |
+
"v_scale": 0.019182477678571428
|
| 45 |
+
},
|
| 46 |
+
"language_model.model.layers.7.self_attn.attn": {
|
| 47 |
+
"k_scale": 0.01787806919642857,
|
| 48 |
+
"v_scale": 0.026241629464285716
|
| 49 |
+
}
|
| 50 |
+
}
|