Daniel

Salt life | Florida Fishing | Still have no idea what twitters for...

[Daniel @Daniel00Higgins]

@_ryu15_ I do this then I try to have them fix the issues and it’s like pulling teeth 🫩

[Daniel @Daniel00Higgins]

@ElBesto37 @RichardSocher What are you talking about they are showing nvfp4. I’m talking quantized to quantized. No one is wasting time running full precision for 1% more accuracy 🤣🤣

[Daniel @Daniel00Higgins]

@RichardSocher This is slower than qwen 27b… is this without mtp or something? Then it would be crazy.

[Daniel @Daniel00Higgins]

@Onepeg Can’t wait to play at 43fps

[Daniel @Daniel00Higgins]

If ComfyUI keeps dying with "access violation" when it swaps GGUF models… try --disable-smart-memory

crashed on me 3 times in one night on my 5070 Ti 16GB
every time right when the next prompt swapped models (GGUF loader partially unloading)

with the flag models fully unload instead
zero crashes the rest of the run

slower swaps, but a queue that doesnt freeze at 2am is worth it

ComfyUI + ComfyUI-GGUF, MiniMax H3 turbo Q4_K_M
https://t.co/xml2IxEaYw

[Daniel @Daniel00Higgins]

@junupark_ Useless for most people as concurrency destroys it.

[Daniel @Daniel00Higgins]

@brad_millscan02 @NousResearch @Teknium @witcheer Yea this shit pisses me off

[Daniel @Daniel00Higgins]

@_ryu15_ You’ve been spotted lol

[Daniel @Daniel00Higgins]

Said last week i couldnt tell Qwen-Image turbo from full… so i actually scored it

same refs + seeds, prompts with stuff you can check (facing direction, pose, props, 2 characters)

turbo 8-step: 98% of checks, 24.8s an image
full 25-step: 100%, 67.2s

1 partial miss for almost 3x the speed. im keeping turbo

int8 DiT, int8 ref cache on GPU, 1024x1024, 2x Arc Pro B60, ComfyUI XPU

[Daniel @Daniel00Higgins]

Qwen-Image 2.1 with 6 reference images was taking 73s an image on my B60s

turned on the int8 ref cache on the GPU
31s first run
25s warm

same job. almost 3x faster

Qwen-Image 2.1 turbo int8, 8 steps, 1024x1024
Qwen3-VL-8B int8 encoder
QwenImage21Cache device=gpu dtype=int8
2x Arc Pro B60 24GB, ComfyUI XPU, ~12.7GB peak per card

https://t.co/GMVMe1QPFs

[Daniel @Daniel00Higgins]

Yea you dont need a BF16 MTP head

swapped my int4 draft head for the original BF16 weights
acceptance at 8k: 51.7% int4 vs 49.5% BF16
at 200k: 62.1% vs 60.1%
and BF16 was 4-6% slower

rerunning the same int4 head moves it ~2.5 points, so its a wash. keep it int4

Qwen3.8 27B GPTQ int4 (Swift 1.5), MTP 6 drafts, 4 users
2x Arc Pro B60, vLLM 0.30 XPU + k8v4 cache
BF16 MTP weights from greglechin on HF

https://t.co/m5SQH6vb96

[Daniel @Daniel00Higgins]

Spent last night trying to beat my own attention kernel on the B60s

fancy idea: pass partial results through L2 in a ring instead of writing them to VRAM
result: 2.3x SLOWER 😩

boring idea: just stop writing the dead rows
21MB less VRAM writes per request, a little faster, still bit-exact

Qwen3.8 27B int4, 200k ctx, 4 users, 2x Arc Pro B60
vLLM XPU + my k8v4 cache

https://t.co/m5SQH6vb96

[Daniel @Daniel00Higgins]

Not specifically Intel but any of the large repos they have too many useless pr’s to filter through and nothing ever gets accepted. I’m glad your stuff did get accepted. I’m tired of everyone making their own tiny projects for small performance stuff instead of the big projects getting it 💪🏽

[Electric Petrolhead - Andy @electricpetrol]

@SawyerMerritt Massive downgrade looks wise compared to the Roadster concept 😭

[Daniel @Daniel00Higgins]

@timosterhus @thsottiaux ^^

[Daniel @Daniel00Higgins]

@julianharris @LottoLabs Never they do it so number big ��

[Daniel @Daniel00Higgins]

@TolkienFragment @LottoLabs Kinda is an understatement.

[Daniel @Daniel00Higgins]

Anyone try the open source rust based Adobe replacements?

https://t.co/psSurtxRxy

[Daniel @Daniel00Higgins]

@CallMeAidyn @EpicGames Claude please add multithreaded async uninstaller.