Park Soobin
subin
AI & ML interests
LLM inference and serving — vLLM internals, chat templates, tokenization, and NPU/GPU deployment.
Recent Activity
new activity about 20 hours ago
vllm-sr/d3-flash:d3-flash on multiple GPUs new activity about 21 hours ago
vllm-sr/vela3-mini:Improving Multilingual Support and NER AccuracyOrganizations
d3-flash on multiple GPUs
2
#5 opened about 20 hours ago
by 
FaisaI
Improving Multilingual Support and NER Accuracy
1
#3 opened about 22 hours ago
by 
BuiDoan
Reject choice questions without criteria on vela3-fast and drop the score claim
1
#2 opened 1 day ago
by 
subin
Return invalid_question for malformed span thresholds and instructions
1
#2 opened 1 day ago
by subin
Return invalid_question for malformed span thresholds and instructions
1
#3 opened 1 day ago
by subin
Return invalid_question for malformed span thresholds and instructions
1
#2 opened 1 day ago
by subin
Return invalid_question for malformed span thresholds and instructions
1
#3 opened 1 day ago
by subin
Keep each thread's adapter gate to itself
1
#1 opened 1 day ago
by subin
Keep each thread's adapter gate to itself
1
#2 opened 1 day ago
by subin
Keep each thread's adapter gate to itself
1
#1 opened 1 day ago
by subin
Keep each thread's adapter gate to itself
1
#2 opened 1 day ago
by subin
Attend without dense masks on long rows on GPUs
#1 opened 2 days ago
by subin
Keep adapter passes out of d3's HIP graphs
#1 opened 2 days ago
by subin
Keep adapter passes out of d3's HIP graphs
#1 opened 2 days ago
by subin
Card: serve with --max-model-len 32768 in Run with vLLM
#7 opened 8 days ago
by subin
Quantized version for vLLM support
👀 1
9
#2 opened 17 days ago
by BuiDoan
Drop the sentence that says the package ships no Python runtime
#2 opened 8 days ago
by subin
Report speed and tasksource agreement per kind instead of pooled
#1 opened 8 days ago
by subin
Card: link the merged record, add quantization tags
1
#1 opened 12 days ago
by subin
how to deploy the model?
5
#1 opened 18 days ago
by 
wanghaikuan