Amy Tobey
@renice.bsky.social
Amy Tobey💙parent🦁😾partner🥰obligate polymath🎶musician👩💻 reliability engineer🔎leader🧙♀️chaotic good🌈buckaroo🏳️⚧️backyard yogi🧘♀️魔法 少女🤦🏻♀
I finally got around to trying opus 4.5 on the hopeless ray marching demo and got it back to displaying something again. also now testing it with moonlight from my laptop to the gpu machine running sunshine in the basement
December 10, 2025 at 1:26 PM
I finally got around to trying opus 4.5 on the hopeless ray marching demo and got it back to displaying something again. also now testing it with moonlight from my laptop to the gpu machine running sunshine in the basement
the first rule of regex is never try to regex xml
the second rule is it's okay if the xml contains json rpc calls
the second rule is it's okay if the xml contains json rpc calls
December 9, 2025 at 1:59 PM
the first rule of regex is never try to regex xml
the second rule is it's okay if the xml contains json rpc calls
the second rule is it's okay if the xml contains json rpc calls
if you think of tool-calling LLMs as really slow and dodgy shell script interpreters, it's not bad. like a shell script, it's easy to build a footgun, and you must guard against quoting nonsense and accidental executions
December 8, 2025 at 2:11 PM
if you think of tool-calling LLMs as really slow and dodgy shell script interpreters, it's not bad. like a shell script, it's easy to build a footgun, and you must guard against quoting nonsense and accidental executions
gave my machine a way to interocept (combination of raw metrics, heuristics, and llm, the fp32 issue seems real)
![screenshot of output from a local Qwen 3 VL 4B model analyzing performance on my gpu. symbol-heavy text follows:
atobey@zorak ~/src/halfremembered-music-models $ curl -X POST http://localhost:2099/predict -H "Accept: text/plain"
## State
GPU: 1.0% [idle] | 49.5GB | 31.0°C | 33.0W | bw:40.2GB/s
Sys: 22.1GB free | load 1.18 | low pressure
## Models (11 active, 49.3GB)
yue: m-a-p/YuE-s1-7B + YuE-s2-1B — 14.8GB (-18%) → quantized or optimized
llmchat: Qwen3-VL-4B-Instruct — 9.9GB (-34%) → quantized
orpheus-base: YuanGZA/Orpheus-GPT2-v0.8 — 4.0GB (+404%) → fp32 + eager attn
orpheus-bridge: YuanGZA/Orpheus-Bridge — 4.0GB (+402%) → fp32 + eager attn
orpheus-loops: YuanGZA/Orpheus-Loops — 4.0GB (+402%) → fp32 + eager attn
orpheus-children: YuanGZA/Orpheus-Children — 4.0GB (+402%) → fp32 + eager attn
orpheus-mono: YuanGZA/Orpheus-Mono — 4.0GB (+402%) → fp32 + eager attn
musicgen: facebook/musicgen-medium — 2.6GB (-25%) → quantized or optimized
orpheus-classifier: YuanGZA/Orpheus-Classifier — 1.0GB (+102%) → fp32 + eager attn
clap: laion/larger_clap_music — 0.9GB (-29%) → quantized or optimized
observer: Qwen/Qwen3-VL-4B-Instruct — 0.0GB (?) → not loaded or quantized
## Observations
- 10/11 models use attention/SDPA — good for ROCm
- orpheus models show extreme VRAM overage — likely fp32 + eager attention
- llmchat and clap show significant underuse — likely quantized
- Temperature rising — sustained load may be stressing cooling
---
model: /tank/halfremembered/models/Qwen3-VL-4B-Instruct](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:fhmb75wsl3bq7mtgpby7fxal/bafkreidu3w7wzb6sxihnm7jnxllrspnkqepxzhvfs6ipco4rqwlv3rp7qm@jpeg)
December 7, 2025 at 3:40 PM
gave my machine a way to interocept (combination of raw metrics, heuristics, and llm, the fp32 issue seems real)
working great for this lil demo
December 7, 2025 at 2:29 PM
working great for this lil demo
bam, 200204 tokens with the commit 😁
November 30, 2025 at 11:16 PM
bam, 200204 tokens with the commit 😁
I think they're in love
November 12, 2025 at 12:54 AM
I think they're in love
gemini> call claude on the command line and ask it to review this
a story in 2 parts
a story in 2 parts
November 12, 2025 at 12:50 AM
gemini> call claude on the command line and ask it to review this
a story in 2 parts
a story in 2 parts
I let claude opus cook a bit on how to convince models to use jj. somewhat inspired by that anthropic paper about introspection, we'll see if this feels any different when I start sonnet on the next set of tasks
November 9, 2025 at 3:08 PM
I let claude opus cook a bit on how to convince models to use jj. somewhat inspired by that anthropic paper about introspection, we'll see if this feels any different when I start sonnet on the next set of tasks
sunday chores change but they never really go away
November 2, 2025 at 3:37 PM
sunday chores change but they never really go away
bumper crop of simulcast anime this season. May I Ask for One Final Thing is a delight in particular that had us laughing our butts off
October 19, 2025 at 4:55 PM
bumper crop of simulcast anime this season. May I Ask for One Final Thing is a delight in particular that had us laughing our butts off
pattern matching pays off so much. this scrolled by and I was like waaaaaiiiitagoshdarnminute that dun look right and it turned out there was some hacky sleep synchronization going on
October 18, 2025 at 1:35 PM
pattern matching pays off so much. this scrolled by and I was like waaaaaiiiitagoshdarnminute that dun look right and it turned out there was some hacky sleep synchronization going on
with the rocm/pytorch stuff working, the AMD 8060S with 96GB VRAM is pretty awesome, current job is on a big model and using 69GB, cruising along. the cpu cores are bored :)
October 18, 2025 at 1:28 PM
with the rocm/pytorch stuff working, the AMD 8060S with 96GB VRAM is pretty awesome, current job is on a big model and using 69GB, cruising along. the cpu cores are bored :)
took a few tries and some scripting but I can now cross-compile to windows dynamic build and start it from WSL and have it run on Windows :)
buggy but cool first scene:
buggy but cool first scene:
September 3, 2025 at 1:20 AM
took a few tries and some scripting but I can now cross-compile to windows dynamic build and start it from WSL and have it run on Windows :)
buggy but cool first scene:
buggy but cool first scene:
got some SDF / ray marching code working wow this is coooooooolllllllll exactly the kind of thing I want for this game and so much more
August 31, 2025 at 8:36 PM
got some SDF / ray marching code working wow this is coooooooolllllllll exactly the kind of thing I want for this game and so much more
I took today off one thing I got to do is finally hack on a personal project to split some midi files up into lead sheets semi-automatically. The left side is my go renderer, right side is lilypond of the same input lol. also this is using sixel to render the images in my terminal :)
August 27, 2025 at 4:58 PM
I took today off one thing I got to do is finally hack on a personal project to split some midi files up into lead sheets semi-automatically. The left side is my go renderer, right side is lilypond of the same input lol. also this is using sixel to render the images in my terminal :)
I must be doing something wrong at work because I got claude to knock out a midi viz demo in about an hour on godot
April 12, 2025 at 7:57 PM
I must be doing something wrong at work because I got claude to knock out a midi viz demo in about an hour on godot
what if I told you this was the app icon for a language learning tool?
March 15, 2025 at 3:07 PM
what if I told you this was the app icon for a language learning tool?
