osobhandClaude Opus 4.7
a5cedfb46a
8-gen bench (4 emotions × 2 corpora) at seed=42 against firdhokk Whisper-LV3: target RAVDESS CREMA-D happy happy (0.999) ✓ happy (0.999) ✓ angry neutral (0.92) sad (0.99) fearful happy (0.998) fearful (0.984) ✓ sad angry (0.99) fearful (0.99) CREMA-D 2/4 vs RAVDESS 1/4. Larger / more naturalistic corpus produces more class-pure fearful direction. Neither corpus solves angry or sad — recipe shifts into 'vague expressivity' rather than class-specific corners. Practical: prefer CREMA-D when available; A/B both per emotion if class precision matters. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
75 lines
2.1 KiB
Rust
75 lines
2.1 KiB
Rust
//! CSM-1B safetensors → quantized GGUF converter.
|
|
//!
|
|
//! Builds the artifact a (future) forked `csm_quantized.rs` would consume.
|
|
//! For now: prints the policy decision report, writes the quantized GGUF.
|
|
//!
|
|
//! Usage:
|
|
//! ```
|
|
//! cargo run -p rtx-csm --release --example quantize -- \
|
|
//! --policy q8 --out /tmp/csm-q8.gguf
|
|
//! cargo run -p rtx-csm --release --example quantize -- \
|
|
//! --policy q4km --out /tmp/csm-q4km.gguf
|
|
//! ```
|
|
|
|
use anyhow::Result;
|
|
use clap::{Parser, ValueEnum};
|
|
use rtx_csm::quantize::{QuantPolicy, convert_to_quantized};
|
|
use rtx_csm::{hub, model};
|
|
|
|
#[derive(Debug, Clone, Copy, ValueEnum)]
|
|
enum PolicyChoice {
|
|
/// Q8_0 on backbone projections, F16/F32 on heads/embeds/decoder
|
|
Q8,
|
|
/// Q4_K on backbone projections, Q8_0 on decoder, F16/F32 on heads/embeds
|
|
Q4km,
|
|
}
|
|
|
|
#[derive(Debug, Parser)]
|
|
#[command(name = "csm-quantize")]
|
|
struct Cli {
|
|
/// Path to input safetensors (defaults to HF-cached sesame/csm-1b weights).
|
|
#[arg(long)]
|
|
input: Option<std::path::PathBuf>,
|
|
|
|
/// Output GGUF file.
|
|
#[arg(long)]
|
|
out: std::path::PathBuf,
|
|
|
|
/// Policy preset.
|
|
#[arg(long, value_enum, default_value = "q8")]
|
|
policy: PolicyChoice,
|
|
|
|
/// Print policy decisions only — don't write anything.
|
|
#[arg(long)]
|
|
dry_run: bool,
|
|
}
|
|
|
|
fn main() -> Result<()> {
|
|
tracing_subscriber::fmt().init();
|
|
let cli = Cli::parse();
|
|
|
|
let input_path = match cli.input {
|
|
Some(p) => p,
|
|
None => hub::resolve_csm_weights()?,
|
|
};
|
|
eprintln!("input: {}", input_path.display());
|
|
eprintln!("output: {}", cli.out.display());
|
|
|
|
let policy = match cli.policy {
|
|
PolicyChoice::Q8 => QuantPolicy::q8_safe(),
|
|
PolicyChoice::Q4km => QuantPolicy::q4km_aggressive(),
|
|
};
|
|
|
|
if cli.dry_run {
|
|
let descs = model::dump_safetensors_keys(&input_path)?;
|
|
let report = rtx_csm::quantize::report(&descs, &policy);
|
|
println!("{report}");
|
|
return Ok(());
|
|
}
|
|
|
|
eprintln!("converting (this can take 30-90 seconds for CSM-1B)...");
|
|
let report = convert_to_quantized(&input_path, &cli.out, &policy)?;
|
|
println!("\n{report}");
|
|
Ok(())
|
|
}
|