DEV Community

Sho Tanaka (tsho) profile picture

Sho Tanaka (tsho)

404 bio not found

Joined Joined on  github website
DeepSpeed ZeRO vs PyTorch FSDP2, Measured: 1.8x on NVLink, Barely Finishing on PCIe (Plus How to Actually Get GPUs on GKE)
tsho profile

DeepSpeed ZeRO vs PyTorch FSDP2, Measured: 1.8x on NVLink, Barely Finishing on PCIe (Plus How to Actually Get GPUs on GKE)

#machinelearning #pytorch #deepspeed #gpu
Comments
10 min read
PyTorch 2.14: clamp's Boundary Gradient Silently Changes, bfloat16 Complex Promotion Now Raises, cholesky/qr Removed
tsho profile

PyTorch 2.14: clamp's Boundary Gradient Silently Changes, bfloat16 Complex Promotion Now Raises, cholesky/qr Removed

#machinelearning #pytorch #python
Comments
4 min read
Turning Qwen2.5-0.5B into a JSON API with SFT — 23% 100% on a Free T4
tsho profile

Turning Qwen2.5-0.5B into a JSON API with SFT — 23% 100% on a Free T4

#llm #machinelearning #posttraining #sft
1
Comments
13 min read
pgvector vs LanceDB: A Reproducible Vector Search Benchmark (100k 1536-dim)
tsho profile

pgvector vs LanceDB: A Reproducible Vector Search Benchmark (100k 1536-dim)

#postgres #database #ai #performance
Comments
6 min read
Why the First Turn in a Coding Agent Can Use So Many Input Tokens — and Why That Gets Better Over Time
tsho profile

Why the First Turn in a Coding Agent Can Use So Many Input Tokens — and Why That Gets Better Over Time

#ai #snowflake #claude #openai
7
Comments
7 min read
TPU v6e vs A100 80GB 2 for Gemma 4 31B on vLLM: 21 Benchmarks Show When Each Wins
tsho profile

TPU v6e vs A100 80GB 2 for Gemma 4 31B on vLLM: 21 Benchmarks Show When Each Wins

#llm #benchmark #gpu #tpu
1
Comments
15 min read
loading...