mirror of
https://github.com/hanzoai/ml.git
synced 2026-11-20 22:46:50 +00:00
Three-way per file through the rename: base 39355c6c91, theirs the upstream
merge of main with 35d7ae7c (via 7ef6a694d), ours origin/main.
Lands the barrier thread pool, runtime CPU feature dispatch, aarch64 and x86
repacked quantized kernels (indexed_gemv, gemv_fused_shared_lhs), quantized
embedding on CPU/Metal/CUDA, paged flash-attn kernels, cutile fused MoE, the
in-place op aliasing API, parking_lot storage locks, Metal buffer builders and
debug labels, safetensors 0.8, hf-hub 1.0, tokenizers 0.23.
Ours keeps ROCm, Vulkan, wgpu, the IQ/MXFP4/NVFP4/TQ dtypes, and the i-quant
weight/activation block split in matmul. Our backends ride InplaceOpN and the
quantized embedding; Metal get_rows gains the IQ kernels. ug leaves, as upstream.
Co-authored-by: Hanzo Dev <[email protected]>
DeepSeek V2
DeepSeek V2 an MoE model featuring MLA (Multi-Latent Attention). There is a lite (16B) and a full (236B) model.
- Context length of 32k tokens (Lite model), 128k tokens (full model)
- 64 routed experts (Lite model), 160 routed experts (full model)
Running the example
$ cargo run --example deepseekv2 --release --features metal -- --prompt "Recursive fibonacci code in Rust:" --which lite --sample-len 150
fn fibonacci(n: u32) -> u32 {
if n <= 1 {
return n;
} else {
return fibonacci(n - 1) + fibonacci(n - 2);
}
}
## Fibonacci code in Python:
def fibonacci(n):
if n <= 1:
return n
else:
return fibonacci(n-1) + fibonacci(n-2)
## Fibonacci code in JavaScript:
function fibonacci(n) {
if (n <= 1