The GPU Is a Queue
A day with 135 GPU jobs and 34 failures, none of which needed the GPU to find. The fix is not a faster GPU, it's treating the card as a shared, budgeted resource with a preflight step before anything touches it.
Read✎Blog
Blog
Notizen aus den Projekten von AMARBARO — gemessene Ergebnisse, das Protokoll, unter dem sie gemessen wurden, und die Vorhersagen, die sich als falsch herausstellten.
Neueste
Models and quality
3 Min. Lesezeit
Krea 2 Wins the Background
Krea 2 won the background task 2 to 1 because it followed the empty-space instruction. Z-Image stayed sharper and faster, so the choice depends on the job.
Serie
Image models in a real pipeline
1 Teile, in Leseordnung
Serie
Gates that lie
5 Teile, in Leseordnung
Serie
Token reduction
1 Teile, in Leseordnung
Serie
Mojo on a phone
1 Teile, in Leseordnung
Serie
A 35B MoE on one consumer card
3 Teile, in Leseordnung
Serie
Kernels on a gaming GPU
5 Teile, in LeseordnungSingle posts
Publishing Without Publishing Yourself
361 commits in mojo-baro's history carried a private email address that is invisible in every checkout of the repo, because it lives in commit metadata, not tracked files. The Android repo separately carried 112 MB of committed build output. Neither would have been caught by reading the code.
Read
Keeping the Failures In
BARO_SPEC defaults to on. Setting it on the MoE model changes nothing, because the actual gate in the code is BARO_SPEC and a model flag that is always false for that architecture, and the document says so in the same line rather than in a footnote.
Read
int8 WMMA Is Not Faster
v_wmma_i32_16x16x16_iu8 issues at 1.007x the rate of the bf16 WMMA instruction on gfx1100, both at 133 T-ops/s. An int8 MMQ prefill kernel built on the expected 2x lost to the project's own bf16 kernel, 0.91x at n=1024, and the lane closed negative.
Read
A Model File That Carries Its Own Kernels
Two bakes of this project's models pass a from-the-file rebuild and gate, 64 of 64 tokens, with no checkout and no external fixture. A third, earlier bake of the same champion model claimed the same thing and was lying: its verify path had quietly fallen back to a local fixture instead of reading its own embedded evidence.
Read
The IPC Primitive Is a Latent Tensor
An OS design in which the message between agents is a latent tensor with an identity, and text is one adapter among several. The transport and tiering numbers are real, the fork gate is resolved, and the thesis experiment has not passed. That experiment asks whether an untrained model reads its own hidden states as reasoning.
ReadEverything, newest first
22 posts- Krea 2 Wins the Background Models and quality 3 Min.
- The Tool Gate Must Own the First Move Measurement 3 Min.
- The Cache Is Not the Only Token Budget Practice 2 Min.
- A Reboot Is Not a Health Check Practice 3 Min.
- Your Reference Fails Its Own Reference Measurement 6 Min.
- The GPU Is a Queue Practice 7 Min.
- Publishing Without Publishing Yourself Practice 8 Min.
- An Arm64 Engine, Bit-Exact on Two Phones, and One Steady on One Edge and mobile 8 Min.
- Keeping the Failures In Practice 7 Min.
- int8 WMMA Is Not Faster GPU kernels 5 Min.
- Compare Bytes, Not Ids Measurement 7 Min.
- A Model File That Carries Its Own Kernels Systems design 9 Min.
- Prefill as a Batch GPU kernels 9 Min.
- Experts in Host RAM Systems design 7 Min.
- Forty-Two to One Hundred Eleven GPU kernels 6 Min.
- The IPC Primitive Is a Latent Tensor Systems design 9 Min.
- A Gate the Candidate Can Write Measurement 7 Min.
- Beating llama.cpp at Decode, Without Speculating GPU kernels 5 Min.
- A Stolen LDS Slot Produces Wrong Tokens, Not an Error GPU kernels 8 Min.
- The Shape That Actually Matters GPU kernels 6 Min.
- Spread Does Not Detect an Inert Parameter Measurement 7 Min.
- Beating hipBLASLt on a Gaming GPU GPU kernels 8 Min.