Add CUDA GPU backend for NVIDIA acceleration

[0xSufi]

Summary

• Add CUDA/cuBLAS GPU backend for NVIDIA GPUs (SM 8.0+), bringing inference from ~40 sec to ~1 sec on RTX 3090
• Implements BF16 weight caching, custom CUDA kernels (attention, RoPE, norms, activations), and monolithic GPU step functions that run all transformer layers with a single CPU↔GPU sync
• New make cuda build target; CUDA and Metal are mutually exclusive at compile time

Performance (RTX 3090, 3.6s test audio)

	Encoder	Decoder (per step)	Decoder (total)
CPU (OpenBLAS)	10,871 ms	~500 ms	28,446 ms
CUDA	225 ms (48x)	12.3 ms (40x)	782 ms (36x)

Architecture

• cuBLAS matmul: BF16 weights cached on GPU after first use, F32→BF16 activation conversion on-device, TF32 tensor ops on Ampere+
• Custom CUDA kernels: rms_norm (shared-memory reduction), silu, gelu, add/mul, apply_rope (paired rotation), causal_attention (online softmax, GQA, sliding
  window), ada_scale, bias_add
• Monolithic step functions: Decoder (26 layers + logits) and encoder (32 layers + final norm) execute entirely on GPU with one cudaStreamSynchronize — no per-layer CPU
  round-trips
• Unified memory KV cache: cudaMallocManaged for zero-copy CPU↔GPU KV caches; GPU writes KV entries via device→managed copy to avoid page thrashing
• Persistent GPU buffers: Single cudaMalloc per component (decoder/encoder), pointer arithmetic for sub-buffers, reused across tokens

New files

• voxtral_cuda.h — C header with CUDA backend API
• voxtral_cuda.cu — Full CUDA implementation (~1100 lines)

Modified files

• Makefile — make cuda target (nvcc + gcc, links cudart/cublas)
• main.c — CUDA init/shutdown lifecycle
• voxtral.c — BF16 + F32 weight warmup, unified memory KV cache allocation
• voxtral_kernels.c — #ifdef USE_CUDA dispatch for bf16 matmul functions
• voxtral_decoder.c — CUDA monolithic decoder step dispatch, unified memory KV cache
• voxtral_encoder.c — CUDA monolithic encoder step dispatch, unified memory KV cache

Test plan

• make cuda builds without errors
• Short audio transcription correct ("Hello, this is a test...")
• Long audio transcription works (JFK 11s, MLK 3min)
• Falls back to CPU path gracefully if CUDA step returns -1
• Verify on different SM 8.x GPUs (tested on SM 8.6 / RTX 3090)

🤖 Generated with Claude Code

[0xSufi]

Ciao Salvatore, mi sembra che Opus 4.6 abbia fatto un buon lavoro qui, magari troppo Linux specific?
Le performances non sono male.

Sei un grande!

[Danmoreng]

How is this different from the existing CUDA PR here #7 ?

[0xSufi]

How is this different from the existing CUDA PR here #7 ?

Well, first of all look at the lines of code this PR added vs the other one, they are substantially different. Also this PR was created entirely with Claude Code (Opus 4.6) admittedly, I am not sure about #7. I have skimmed it briefly and it looks like it's Windows specific vs this PR, which should be Linux specific (I asked antirez about it).

Edit: Here's Opus 4.6 answer to your question:

● Both PRs add CUDA GPU backends for NVIDIA acceleration to voxtral.c, but they differ significantly in scope and approach:

  PR #12 — "Add CUDA GPU backend for NVIDIA acceleration"

  - Size: ~1,436 additions, 8 files changed
  - Scope: Focused, clean CUDA backend
  - Implementation: Single .cu file (~1100 lines), cuBLAS for matmul, custom kernels (RMSNorm, SiLU, GeLU, RoPE, causal attention, etc.)
  - Architecture: Monolithic GPU step functions (one cudaStreamSynchronize per step), unified memory (cudaMallocManaged) for KV caches, BF16 weight caching
  - Performance: RTX 3090, 3.6s audio — encoder 48x faster, decoder 40x faster (~1s total vs ~40s CPU)
  - Platform: Linux only
  - State: Open

  PR #7 — "CUDA backend (NVIDIA/WSL2) + faster-than-real-time STT"

  - Size: ~12,156 additions, 30 files changed
  - Scope: Much larger — CUDA backend + Windows support + many optimizations
  - Implementation: Separate voxtral_cuda.c (~6382 lines) + voxtral_cuda_kernels.cu (~2927 lines), uses CUDA Driver API + cuBLAS/cuBLASLt
  - Extra features:
    - Many tuning knobs (VOX_CUDA_FAST, VOX_CUDA_PIPELINE_FULL, VOX_CUDA_LOGITS_INT8, cuBLASLt autotune, etc.)
    - Fused kernels (RMSNorm + ada_scale + BF16 cast in one kernel)
    - CUDA graphs for decoder loop
    - INT8-quantized LM head (optional)
    - Pre-compiled cubin blob (avoids PTX JIT issues on WSL2)
    - Native Windows build (PowerShell scripts, WASAPI microphone capture)
  - Performance: RTX 3080 Ti, 180s audio — 5.5x real-time with CUDA fast + INT8
  - Platform: Linux, WSL2, and Windows
  - State: Open

  Key Differences

  ┌────────────────────┬─────────────────┬────────────────────────────────────┐
  │       Aspect       │     PR #12      │               PR #7                │
  ├────────────────────┼─────────────────┼────────────────────────────────────┤
  │ Lines added        │ ~1,400          │ ~12,000                            │
  ├────────────────────┼─────────────────┼────────────────────────────────────┤
  │ Files changed      │ 8               │ 30                                 │
  ├────────────────────┼─────────────────┼────────────────────────────────────┤
  │ CUDA API           │ Runtime API     │ Driver API + cuBLASLt              │
  ├────────────────────┼─────────────────┼────────────────────────────────────┤
  │ KV cache           │ Unified memory  │ Device memory + lazy host download │
  ├────────────────────┼─────────────────┼────────────────────────────────────┤
  │ Windows support    │ No              │ Yes (WASAPI mic, PowerShell build) │
  ├────────────────────┼─────────────────┼────────────────────────────────────┤
  │ Optimization knobs │ Minimal         │ Extensive (graphs, INT8, autotune) │
  ├────────────────────┼─────────────────┼────────────────────────────────────┤
  │ Complexity         │ Simple, focused │ Feature-rich, many options         │
  └────────────────────┴─────────────────┴────────────────────────────────────┘

  In short: PR #12 is a clean, minimal CUDA backend; PR #7 is a comprehensive, heavily optimized CUDA backend with Windows support and many tuning options.