gahow/xtrain
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
19 Commits 1 Branch 0 Tags
603c85e1e0818173e7a092a8eb93ef05d98c82da
Commit Graph

4 Commits

Author SHA1 Message Date
Gahow Wang
e3912c2380 model: tiny RoPE+RMSNorm+SwiGLU transformer + overfit test 
New crate xtrain-model: a from-scratch decoder built entirely from the
autodiff op set.
- Config (tiny: dim=32, 2 layers, 2 heads, head_dim=16, ffn=64).
- TinyTransformer: embedding -> N x {pre-RMSNorm -> multi-head causal
  attention (RoPE, additive causal mask, per-head SDPA) -> residual;
  pre-RMSNorm -> SwiGLU MLP -> residual} -> final RMSNorm -> LM head.
  x@W weight convention (engine GEMM is plain A@B); dim=n_heads*head_dim.
- params()/zero_grad-able leaves for the optimizer; param_to_host export.
- overfit test: char-level bring-up (embedded text -> vocab -> shifted
  targets), minimal hand-written GD (p -= lr*grad) memorises one fixed
  batch -> loss ~0 + greedy argmax matches targets. End-to-end fwd+bwd
  correctness signal. Gated #![cfg(not(no_cuda))].

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
 2026-06-15 16:05:20 +08:00
Gahow Wang
9ca98efd98 autodiff: finite-diff gradient-check harness 
New xtrain-autodiff crate with a reusable central finite-difference
gradient check: grad_check(x, shape, f, analytic_grad, cfg) compares an
analytic gradient against (f(x+ε)-f(x-ε))/2ε per element with a relative
tolerance. Host-only (no CUDA): the loss closure owns any GPU work, so
T4's per-op backward checks can reuse it directly. Includes host unit
tests (sum(x²) grad 2x passes; a wrong grad is rejected).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
 2026-06-15 15:26:42 +08:00
Gahow Wang
fbd07a578c tensor: minimal Tensor crate over xtrain-cuda 
New xtrain-tensor crate: DType (F32), shape/stride helpers, Arc-counted
host/device Storage with CPU↔CUDA copy, and a contiguous Tensor with
creation, host↔device transfer, and a scale() op driving the elementwise
kernel. GPU integration tests (host↔device roundtrip + scale correctness)
gated behind not(no_cuda); a thin build.rs emits the no_cuda cfg so the
kernel call sites compile out locally.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
 2026-06-15 15:13:06 +08:00
Gahow Wang
92acf9f413 T1: scaffold repo + Rust/CUDA build chain (vecadd smoke test) 
Stand up the xtrain project skeleton: a Cargo workspace mirroring xserv's
csrc/ + crates/ layout, with a single xtrain-cuda crate that wraps the CUDA
Runtime over hand-written extern "C" FFI. build.rs compiles csrc/test/vecadd.cu
via the cc crate targeting sm_120 (RTX 5090) and links cudart.

A gated integration test runs the vector-add kernel on the GPU and asserts the
result. When nvcc is absent (local GPU-less machine), build.rs skips CUDA
compilation and sets a `no_cuda` cfg so host-side cargo check still works.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
 2026-06-15 14:42:43 +08:00