tinygrad · Issues· 193 open

NV eGPU (TinyGPU, RTX 4090): tinygrad.llm decodes at ~1 % of VRAM bandwidth by default; JITBEAM=2 gives 9× on Q8_0 but a BEAM candidate faulted the device

v0.14.0 wheel is missing tinygrad/llm/kernels — `python -m tinygrad.llm` fails at import on a pip install

M4 Max 36GB / 7900GRE (RDNA3) / ADT-UT3G

hcq2 + chestnut broken when `GMMU=1`

IMAGE mode changes integer matmul dtype and loses precision

CPU:LVP float64-to-float16 casts misround values just beyond a half midpoint

QCOM disassembly passes a packed chip identifier instead of the decoder generation

image_dot indexes a nonexistent axis for a right-hand vector

BEAM=2 hangs the GPU on gfx1201 (RDNA4) with bf16 matmul, requiring driver reset

AMD RDNA4 over USB (ASM2464 chestnut + RX 9060 8GB): JIT graph replay hangs; page-table assertion under reduced context

Lazy-load GGUF MoE experts from DISK on demand

[macOS TinyGPU] RTX 5090 Laptop (10de:2c18) fails MAP_BAR for BAR1 via FEVM FNGT5 PRO

Q8_0 Qwen 27B decode is ~85x below isolated GEMV bandwidth

can we build this inside the linairizer

NV on Intel Mac Pro 2019 with RTX 3060 in internal PCIe slot (no Thunderbolt) fails with GSP halt before RPC queue and no panic, plus dext caps DMA at 32 segments