Files
rustytorch/crates/models/rtx-csm/src/sampler.rs
T
osobhandClaude Opus 4.7 15dd3575d4 Add rtx-csm: Rust-native port of Sesame CSM-1B with LoRA voice cloning
A new model crate at crates/models/rtx-csm implementing end-to-end
inference, quantization, and fine-tuning for Sesame's Conversational
Speech Model (CSM-1B). Built on candle 0.9 + Kyutai Mimi codec.

Key capabilities:
- Inference (FP F16 on Metal, F32 on CPU, BF16 on CUDA)
- Quantized inference (Q8_0 / Q4_K_M GGUF, ~3x speedup, ~50% memory)
- Streaming Mimi decode with proper StreamTensor state machine
- In-context voice cloning via SpeakerProfile
- Classifier-Free Guidance (Koel-TTS recipe)
- Long-form chunked generation with rolling context
- Audio post-processing (HPF + declick + EBU R128 LUFS)
- Text input normalization (brackets, times, unicode, length caps)
- Frame-level repetition guard (loop-escape)
- Top-k + top-p sampling
- LoRA fine-tuning end-to-end (training + inference, on FP and Q8 bases)
- In-process Whisper ASR via whisper-rs (under --features asr)
- Standalone TTS HTTP server (Axum)
- Bench harness with manifest export + per-prompt WER

Phases delivered: quantization, ASR/WER eval, LoRA voice cloning, HTTP
service. AudioSeal/WavLM/Unmute remain as documented future work.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
2026-04-25 18:33:57 -07:00

49 lines
1.4 KiB
Rust

use crate::error::Result;
use candle_core::Tensor;
use candle_transformers::generation::{LogitsProcessor, Sampling};
pub const DEFAULT_TEMPERATURE: f64 = 0.9;
pub const DEFAULT_TOPK: usize = 50;
pub const DEFAULT_TOPP: f64 = 0.9;
pub struct CsmSampler {
inner: LogitsProcessor,
}
impl CsmSampler {
/// Build a sampler with explicit knobs. `top_p == 0.0` (or `>= 1.0`) disables
/// nucleus filtering and falls back to pure top-k. `temperature <= 0.0` selects
/// argmax (deterministic) regardless of top_k/top_p.
pub fn new(seed: u64, temperature: f64, top_k: usize, top_p: f64) -> Self {
let sampling = if temperature <= 0.0 {
Sampling::ArgMax
} else if top_p > 0.0 && top_p < 1.0 {
Sampling::TopKThenTopP {
k: top_k,
p: top_p,
temperature,
}
} else {
Sampling::TopK {
k: top_k,
temperature,
}
};
Self {
inner: LogitsProcessor::from_sampling(seed, sampling),
}
}
pub fn default_for_csm(seed: u64) -> Self {
Self::new(seed, DEFAULT_TEMPERATURE, DEFAULT_TOPK, DEFAULT_TOPP)
}
pub fn sample(&mut self, logits: &Tensor) -> Result<u32> {
Ok(self.inner.sample(logits)?)
}
pub fn inner_mut(&mut self) -> &mut LogitsProcessor {
&mut self.inner
}
}