A few new CUDA hacker friends joined the effort and now llm.c is only 2X slower than PyTorch (fp32, forward pass) compared to 4 days ago, when it was at 4.2X slower The biggest improvements were:
– turn on TF32 (NVIDIA TensorFLoat-32) instead of FP32 for matmuls. This is a
LLMS
-
llm.c Performance Optimization: Closing Gap with PyTorch
By
–
-
Grok-1.5V becomes multimodal with sample imagery analysis
By
–
Grok goes multimodal with Grok-1.5V Sample imagery analysis:
-
v1.6.5 release: GPT-4 Turbo, Gemini 1.5, Command-R support
By
–
Quick minor release v1.6.5, supporting: GPT 4-Turbo Gemini 1.5-pro-latest Cohere Command-R We'll be adding support for images/videos upload for Gemini, and function calling with Anthropic Claude3 soon! Back to building
-
Bixtral Fine-Tuned Model Approaches Claude Opus Performance
By
–
Bixtral is a beast. A well-done fine-tune doesn’t feel far off from Opus.
-
Closing the Performance Gap in AI Systems
By
–
yeah true I kind of am lmao but there’s a Twitter thread where ppl are trying to close the performance gap
-
LLM Gullibility Undermines Agent System Viability
By
–
LLMs are inherently gullible. Until we solve that a lot of these "agent" dreams aren't actually going to work very well in practice!
-
DBRX Model Assessment: Early Judgment and Performance Catchup
By
–
judging dbrx too quickly? its big but we’ll catch up to it rather than other way around
-
1-bit LLMs: Optimizing Parameters and Data with Chinchilla Laws
By
–
My favorite part about @honicky
's Paper Club session this week on the 1-bit LLMs paper – relating it to @jefrankle
's Beyond Chinchilla laws and adjusting the equations for the memory/latency characteristics of 1-bit LLMs to derive an optimal param count/data size to aim for. no -
Mixtral 8x22b Instant now available on Poe platform
By
–
You can try it at https://
poe.com/Mixtral8x22b-I
nst-FW
… and across all Poe apps. Enjoy! (2/2) -
Mixtral 8x22B Now Available on Poe via Fireworks AI
By
–
Now on Poe: Mixtral 8x22B! This bot, hosted and fine-tuned by @FireworksAI_HQ
, is among the first instruct-tuned variants of the new Mixtral 8x22B model. (1/2)