Image-Text-to-Text
Transformers
Safetensors
hy_v4_flash
text-generation
nvfp4
fp8
modelopt
4-bit precision
quantized
Mixture of Experts
mixture-of-experts
multimodal
vision
long-context
speculative-decoding
dgx-spark
gb10
hunyuan
hy4-flash
instruct
conversational
8-bit precision
Instructions to use Avifenesh/Hy4-flash-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Avifenesh/Hy4-flash-NVFP4 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="Avifenesh/Hy4-flash-NVFP4") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoModelForSeq2SeqLM model = AutoModelForSeq2SeqLM.from_pretrained("Avifenesh/Hy4-flash-NVFP4", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use Avifenesh/Hy4-flash-NVFP4 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Avifenesh/Hy4-flash-NVFP4" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Avifenesh/Hy4-flash-NVFP4", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/Avifenesh/Hy4-flash-NVFP4
- SGLang
How to use Avifenesh/Hy4-flash-NVFP4 with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "Avifenesh/Hy4-flash-NVFP4" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Avifenesh/Hy4-flash-NVFP4", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "Avifenesh/Hy4-flash-NVFP4" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Avifenesh/Hy4-flash-NVFP4", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use Avifenesh/Hy4-flash-NVFP4 with Docker Model Runner:
docker model run hf.co/Avifenesh/Hy4-flash-NVFP4
Download quant_config.json from Avifenesh/Hy4-flash-NVFP4: direct link, hf CLI and curl.
- Browser
- Download file 42.1 kB
-
https://huggingface.co/Avifenesh/Hy4-flash-NVFP4/resolve/main/quant_config.json
- Command line
-
hf download hf://Avifenesh/Hy4-flash-NVFP4/quant_config.json
-
curl -L -o quant_config.json https://huggingface.co/Avifenesh/Hy4-flash-NVFP4/resolve/main/quant_config.json
42.1 kB
{
"producer": {
"repo": "Avifenesh/Hy4-flash-NVFP4",
"date": "2026-11-18",
"source": "tencent/Hy4-flash@d88c9dfd96ba4f2e6b92adba0b8e8f1c8696f959",
"note": "Modified from tencent/Hy4-flash: weights quantized as listed below; see README.md"
},
"quantization": {
"quant_algo": "MIXED_PRECISION",
"kv_cache_quant_algo": null,
"activations": "not quantized (weight-only)",
"layout": "LAYOUT.md",
"quantized_layers": {
"model.dfly_layers.0.mlp.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.0.mlp.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.0.mlp.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.0.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.0.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.0.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.1.mlp.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.1.mlp.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.1.mlp.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.1.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.1.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.1.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.2.mlp.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.2.mlp.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.2.mlp.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.2.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.2.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.2.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.3.mlp.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.3.mlp.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.3.mlp.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.3.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.3.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.3.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.4.mlp.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.4.mlp.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.4.mlp.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.4.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.4.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.4.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.hidden_correction.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.hidden_correction.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.hidden_correction.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.layers.1.engram_layer.multi_head_embedding.embedding": {
"quant_algo": "FP8_ROW"
},
"model.layers.1.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.1.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.1.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.1.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.1.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.1.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.10.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.10.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.10.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.10.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.10.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.10.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.11.engram_layer.multi_head_embedding.embedding": {
"quant_algo": "FP8_ROW"
},
"model.layers.11.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.11.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.11.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.11.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.11.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.11.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.12.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.12.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.12.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.12.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.12.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.12.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.13.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.13.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.13.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.13.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.13.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.13.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.14.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.14.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.14.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.14.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.14.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.14.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.15.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.15.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.15.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.15.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.15.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.15.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.16.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.16.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.16.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.16.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.16.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.16.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.17.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.17.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.17.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.17.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.17.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.17.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.18.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.18.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.18.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.18.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.18.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.18.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.19.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.19.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.19.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.19.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.19.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.19.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.2.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.2.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.2.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.2.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.2.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.2.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.20.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.20.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.20.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.20.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.20.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.20.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.21.engram_layer.multi_head_embedding.embedding": {
"quant_algo": "FP8_ROW"
},
"model.layers.21.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.21.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.21.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.21.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.21.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.21.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.22.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.22.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.22.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.22.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.22.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.22.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.23.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.23.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.23.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.23.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.23.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.23.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.24.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.24.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.24.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.24.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.24.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.24.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.25.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.25.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.25.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.25.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.25.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.25.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.26.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.26.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.26.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.26.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.26.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.26.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.27.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.27.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.27.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.27.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.27.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.27.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.28.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.28.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.28.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.28.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.28.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.28.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.29.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.29.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.29.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.29.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.29.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.29.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.3.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.3.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.3.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.3.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.3.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.3.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.30.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.30.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.30.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.30.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.30.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.30.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.31.engram_layer.multi_head_embedding.embedding": {
"quant_algo": "FP8_ROW"
},
"model.layers.31.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.31.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.31.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.31.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.31.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.31.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.32.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.32.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.32.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.32.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.32.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.32.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.33.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.33.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.33.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.33.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.33.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.33.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.34.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.34.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.34.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.34.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.34.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.34.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.35.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.35.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.35.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.35.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.35.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.35.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.36.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.36.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.36.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.36.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.36.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.36.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.37.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.37.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.37.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.37.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.37.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.37.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.38.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.38.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.38.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.38.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.38.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.38.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.39.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.39.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.39.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.39.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.39.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.39.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.4.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.4.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.4.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.4.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.4.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.4.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.40.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.40.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.40.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.40.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.40.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.40.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.41.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.41.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.41.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.41.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.41.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.41.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.5.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.5.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.5.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.5.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.5.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.5.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.6.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.6.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.6.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.6.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.6.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.6.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.7.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.7.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.7.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.7.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.7.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.7.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.8.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.8.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.8.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.8.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.8.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.8.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.9.mlp.experts.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.9.mlp.experts.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.9.mlp.experts.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per expert"
},
"model.layers.9.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.9.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.layers.9.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
}
}
}
}
{
"producer": {
"repo": "Avifenesh/Hy4-flash-NVFP4",
"date": "2026-11-18",
"source": "tencent/Hy4-flash@d88c9dfd96ba4f2e6b92adba0b8e8f1c8696f959",
"note": "Modified from tencent/Hy4-flash: weights quantized as listed below; see README.md"
},
"quantization": {
"quant_algo": "MIXED_PRECISION",
"kv_cache_quant_algo": null,
"activations": "not quantized (weight-only)",
"layout": "LAYOUT.md",
"quantized_layers": {
"model.dfly_layers.0.mlp.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.0.mlp.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.0.mlp.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.0.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.0.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.0.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.1.mlp.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.1.mlp.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.1.mlp.up_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.1.self_attn.linear_gate": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.1.self_attn.o_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.1.self_attn.q_b_proj": {
"quant_algo": "FP8_BLOCK",
"block": [
128,
128
]
},
"model.dfly_layers.2.mlp.down_proj": {
"quant_algo": "NVFP4",
"group_size": 16,
"weight_scale": "E4M3 per 16",
"weight_scale_2": "FP32 per tensor"
},
"model.dfly_layers.2.mlp.gate_proj": {
"quant_algo": "NVFP4",
"group_size": 16,