Hugging Face's logo
Activity Feed

AI & ML interests

Canada Quant Labs — Canada's open-weight model lab. We train, quantize, and ship sovereign reference models for regulated industries (legal, medical, defence, finance) on a DGX B300 at Equinix Vancouver. Upstream contributors to vLLM and llm-compressor. Recipes: W4A16, NVFP4, MXFP4. Built in Victoria, BC. partnerships@cql.ca · cql.ca

Recent Activity
pastapaul  new activity 13 days ago
canada-quant/GLM-5.2-W4A16-MTP:Loading the model
pastapaul  updated a model 13 days ago
canada-quant/GLM-5.2-W4A16-MTP
pastapaul  new activity 15 days ago
canada-quant/GLM-5.2-W4A16-MTP:Can you run NVFP4 + Marlin?
View all activity
pastapaul 
in canada-quant/GLM-5.2-W4A16-MTP 13 days ago

Loading the model

4
#1 opened 16 days ago by
madeby561
pastapaul 
updated a model 13 days ago
pastapaul 
in canada-quant/GLM-5.2-W4A16-MTP 15 days ago

Can you run NVFP4 + Marlin?

5
#2 opened 16 days ago by
einsteiner1983
pastapaul 
updated a Space 16 days ago
pastapaul 
published a model 18 days ago
pastapaul 
in canada-quant/DeepSeek-V4-Flash-NVFP4-FP8-MTP about 1 month ago

Docker Image

8
#1 opened about 2 months ago by
mtcl
pastapaul 
in canada-quant/DeepSeek-V4-Flash-W4A16-FP8 about 1 month ago

Compatible version with Ampere? SM8.6

1
#6 opened about 2 months ago by
bullerwins

Small patch to fix MTP on current vLLM (8th of June, 26)

7
#8 opened about 1 month ago by
dnhkng
pastapaul 
in canada-quant/DeepSeek-V4-Flash-W4A16-FP8-MTP about 1 month ago

MTP BF16 draft weights may need explicit quant_config exclusion for vLLM

2
#2 opened about 1 month ago by
yangsiqt2

vLLM compressed-tensors loader fails on shared_experts.gate_up_proj — missing quant target for fused shared-expert projection?

3
#1 opened about 1 month ago by
ajdoosh
pastapaul 
updated a model about 1 month ago
pastapaul 
updated a dataset about 2 months ago
pastapaul 
published a dataset about 2 months ago
pastapaul 
updated a model about 2 months ago
pastapaul 
in canada-quant/DeepSeek-V4-Flash-W4A16-FP8 about 2 months ago

2026-05-26 dual-spark bench: GSM8K 94-98%, AIME 4/5, concurrency wall 12→125 tok/s
#7 opened about 2 months ago by
pastapaul
pastapaul 
in canada-quant/DeepSeek-V4-Pro-NVFP4-FP8-MTP about 2 months ago

Lossless?

👀 1
1
#1 opened about 2 months ago by
lukealonso
pastapaul 
updated a model about 2 months ago
pastapaul 
in canada-quant/DeepSeek-V4-Flash-W4A16-FP8 about 2 months ago

Does the Redhat accuracy loss concern apply here also?

1
#5 opened about 2 months ago by
1anH
pastapaul 
updated 2 models about 2 months ago