TMP

Cargo.toml

@@ -75,10 +75,12 @@ compact_str = "0.9.0"
 dary_heap = "0.3.8"
 fancy-regex = "0.16.2"
 indicatif = { version = "0.18.3", features = ["improved_unicode"] }
+itertools = "0.14.0"
 futures-util = "0.3.31"
 ndarray = { version = "0.16.1", features = ["serde"] }
 parking_lot = "0.12.5"
 parquet = "56.2.0"
+rand = "0.9.2"
 rayon = "1.11.0"
 reqwest = { version = "0.12.24", features = ["json", "stream"] }
 serde = "1.0.228"
@@ -91,7 +93,11 @@ tracing-indicatif = "0.3.13"
 tracing-subscriber = { version = "0.3.20", features = ["env-filter", "json"] }
 url = "2.5.7"
 
+
+burn-train = { git = "https://github.com/tracel-ai/burn.git", default-features = false }
+
 [workspace.dependencies.burn]
-version = "0.19.1"
+#version = "0.19.1"
+git = "https://github.com/tracel-ai/burn.git"
 default-features = false
-features = ["std", "fusion", "ndarray", "webgpu", "cuda"]
+features = ["std", "fusion", "ndarray", "webgpu", "cuda", "autodiff"]

crates/llmfs/Cargo.toml

@@ -10,15 +10,18 @@ workspace = true
 [dependencies]
 tokenizer = { workspace = true }
 
+ahash = { workspace = true }
 anstyle = { workspace = true }
 anyhow = { workspace = true }
 burn = { workspace = true }
+burn-train = { workspace = true }
 clap = { workspace = true }
 futures-util = { workspace = true }
 indicatif = { workspace = true }
 ndarray = { workspace = true }
 parking_lot = { workspace = true }
 parquet = { workspace = true }
+rand = { workspace = true }
 rayon = { workspace = true }
 reqwest = { workspace = true }
 serde = { workspace = true }

crates/llmfs/src/command/mod.rs

@@ -15,7 +15,6 @@ pub enum SubCommand {
 		#[command(flatten)]
 		args: train_tokenizer::TrainTokenizerArgs,
 	},
-
 	/// Sample data
 	SampleData {
 		#[command(flatten)]

crates/llmfs/src/command/sample_data.rs

@@ -1,100 +1,34 @@
+use ahash::AHasher;
 use anyhow::{Context, Result};
 use burn::{
 	Tensor,
-	backend::{Cuda, cuda::CudaDevice},
+	backend::{Autodiff, Cuda, cuda::CudaDevice},
-	module::{Module, Param, ParamId},
+	config::Config,
+	module::{AutodiffModule, Module, Param, ParamId},
 	nn::{
 		Dropout, Embedding, EmbeddingConfig, LayerNorm, LayerNormConfig,
+		loss::CrossEntropyLossConfig,
 		transformer::{PositionWiseFeedForward, PositionWiseFeedForwardConfig},
 	},
-	prelude::Backend,
+	optim::{AdamConfig, GradientsParams, Optimizer},
+	prelude::{Backend, ToElement},
 	tensor::{Bool, Distribution, Int, activation::softmax},
 };
+use burn_train::ClassificationOutput;
 use clap::Args;
 use indicatif::MultiProgress;
-use ndarray::Array2;
+use ndarray::{Array1, Array2};
-use std::{f32, fs::File, path::PathBuf};
+use std::{
-use tokenizer::Tokenizer;
+	collections::VecDeque,
-
+	f32,
-use crate::data_reader::DataReader;
+	fs::File,
-
+	hash::Hasher,
-#[derive(Debug, Args, Clone)]
+	path::{Path, PathBuf},
-
-pub struct SampleDataArgs {
-	/// Path to training data
-	#[clap(long, default_value = "data")]
-	data_dir: PathBuf,
-
-	/// Path to tokenizer
-	#[clap(long)]
-	tokenizer: PathBuf,
-
-	/// How many texts to return
-	#[clap(long, short = 'n', default_value = "10")]
-	n: usize,
-
-	/// How many texts to skip
-	#[clap(long, short = 's', default_value = "0")]
-	skip: usize,
-}
-
-#[derive(Debug, Clone)]
-pub struct Config {
-	/// Number of tokens
-	pub vocab_size: u32,
-
-	/// Maximum number of input tokens with positional embeddings
-	pub context_size: usize,
-
-	/// Dimension of each token's embedding
-	pub embed_dim: usize,
-
-	/// Number of attention heads
-	pub n_heads: usize,
-
-	/// Dimension of each attn head
-	pub head_dim: usize,
-
-	/// Number of transformer blocks
-	pub n_layers: usize,
-
-	pub embed_drop_rate: f64,
-	pub attention_drop_rate: f64,
-	pub shortcut_drop_rate: f64,
-}
-
-impl SampleDataArgs {
-	pub fn run(self, _mp: Option<MultiProgress>) -> Result<()> {
-		let device = CudaDevice::new(0);
-
-		let iter = DataReader::new(1, &self.data_dir).context("while initializing data reader")?;
-
-		let tokenizer = File::open(&self.tokenizer).context("while opening tokenizer")?;
-		let tokenizer: Tokenizer =
-			serde_json::from_reader(tokenizer).context("while loading tokenizer")?;
-
-		let config = Config {
-			vocab_size: tokenizer.vocab_size(),
-			context_size: 4,
-			embed_dim: 768,
-			n_heads: 12,
-			head_dim: 64, // = 768 / 12
-			n_layers: 12,
-			embed_drop_rate: 0.1,
-			attention_drop_rate: 0.1,
-			shortcut_drop_rate: 0.1,
 };
+use tokenizer::Tokenizer;
+use tracing::{debug, info};
 
-		let stride = config.context_size;
+use crate::data_reader::{DataReader, DataReaderError};
-
-		let batch_size = 10;
-		let mut input_batch = Vec::with_capacity(batch_size);
-		let mut output_batch = Vec::with_capacity(batch_size);
-
-		#[expect(clippy::unwrap_used)] // Lazy error handling
-		let iter = iter.map(|x| x.unwrap()).skip(self.skip).take(self.n);
-
-		let model = GptModel::new(&config, &device);
 
 // Text generation routine
 
@@ -134,87 +68,289 @@ impl SampleDataArgs {
 }
 */
 
-		for i in iter {
+struct TrainTestIterator<'a, B: Backend> {
-			let tokens = tokenizer.encode(&i);
+	reader: DataReader,
 
-			// Skip small texts.
+	ccfg: &'a ComputeConfig,
+	mcfg: &'a GptModelConfig,
+	tokenizer: &'a Tokenizer,
+	eval: bool,
+	device: &'a B::Device,
+
+	error: bool,
+
+	// Tokenized input/output pairs
+	pairs: VecDeque<(Vec<u32>, u32)>,
+}
+
+impl<'a, B: Backend> TrainTestIterator<'a, B> {
+	pub fn new(
+		data_dir: impl AsRef<Path>,
+		ccfg: &'a ComputeConfig,
+		mcfg: &'a GptModelConfig,
+		tokenizer: &'a Tokenizer,
+		eval: bool,
+		device: &'a B::Device,
+	) -> Result<Self, std::io::Error> {
+		let reader = DataReader::new(10, data_dir)?;
+
+		Ok(Self {
+			reader,
+			ccfg,
+			mcfg,
+			tokenizer,
+			eval,
+			device,
+
+			error: false,
+			pairs: VecDeque::new(),
+		})
+	}
+}
+
+impl<B: Backend> Iterator for TrainTestIterator<'_, B> {
+	type Item = Result<TrainBatch<B>, DataReaderError>;
+
+	fn next(&mut self) -> Option<Self::Item> {
+		if self.error {
+			return None;
+		}
+
+		let mut inputs = Vec::with_capacity(self.ccfg.batch_size);
+		let mut targets = Vec::with_capacity(self.ccfg.batch_size);
+		let stride = self.mcfg.context_size;
+
+		while inputs.len() < self.ccfg.batch_size {
+			match self.pairs.pop_front() {
+				Some((i, t)) => {
+					// train/test split
+					{
+						let mut hasher = AHasher::default();
+						hasher.write(self.ccfg.eval_salt.as_bytes());
+
+						// Don't care about endianness, ahash output is unstable anyway
+						hasher.write(unsafe { std::mem::transmute(&i[..]) });
+						hasher.write_u32(t);
+
+						let test = // is this point in the test set?
+							hasher.finish() > (u64::MAX as f64 * self.ccfg.eval_frac).to_u64();
+
+						if test ^ self.eval {
+							continue;
+						}
+					}
+
+					inputs.push(i);
+					targets.push(t);
+				}
+
+				None => {
+					let text = match self.reader.next() {
+						None => break,
+						Some(Ok(x)) => x,
+						Some(Err(x)) => {
+							self.error = true;
+							return Some(Err(x));
+						}
+					};
+
+					let emb = self.tokenizer.encode(&text);
+
+					// Skip small texts
+					//
 					// TODO: do this better
+					// TODO: maybe using <|bos|>?
 					// TODO: non-uniform batches?
-			if tokens.len() < config.context_size {
+					if emb.len() < self.mcfg.context_size {
 						continue;
 					}
 
-			for (a, b) in tokens.windows(config.context_size).step_by(stride).zip(
+					let pairs = emb
-				tokens[stride..]
+						.windows(self.mcfg.context_size + 1)
-					.windows(config.context_size)
+						.step_by(stride)
-					.step_by(stride),
+						.map(|x| {
-			) {
+							(
-				input_batch.push(a.to_owned());
+								x[..self.mcfg.context_size].to_vec(),
-				output_batch.push(b.to_owned());
+								x[self.mcfg.context_size],
+							)
+						});
+
+					self.pairs.extend(pairs);
+				}
+			}
+		}
+
+		if inputs.is_empty() {
+			return None;
+		}
+
+		let shape = [inputs.len(), self.mcfg.context_size];
+
+		// Arrange data in memory
+		let inputs: Array2<u32> = Array2::from_shape_fn(shape, |(a, b)| inputs[a][b]);
+		let targets: Array1<u32> = Array1::from_vec(targets);
+
+		// Create tensors on gpu
+		#[expect(clippy::unwrap_used)]
+		let inputs =
+			Tensor::<B, 1, Int>::from_ints(inputs.as_slice().unwrap(), self.device).reshape(shape);
+
+		#[expect(clippy::unwrap_used)]
+		let targets = Tensor::<B, 1, Int>::from_ints(targets.as_slice().unwrap(), self.device);
+
+		return Some(Ok(TrainBatch { inputs, targets }));
+	}
+}
+
+#[derive(Debug, Args, Clone)]
+
+pub struct SampleDataArgs {
+	/// Path to training data
+	#[clap(long, default_value = "data")]
+	data_dir: PathBuf,
+
+	/// Path to tokenizer
+	#[clap(long)]
+	tokenizer: PathBuf,
+
+	/// How many texts to return
+	#[clap(long, short = 'n', default_value = "10")]
+	n: usize,
+}
+
+pub struct ComputeConfig {
+	pub batch_size: usize,
+	pub eval_frac: f64,
+	pub eval_salt: String,
+}
+
+impl SampleDataArgs {
+	pub fn run(self, _mp: Option<MultiProgress>) -> Result<()> {
+		let device = CudaDevice::new(0);
+		//let device = WgpuDevice::DiscreteGpu(0);
+
+		let tokenizer = File::open(&self.tokenizer).context("while opening tokenizer")?;
+		let tokenizer: Tokenizer =
+			serde_json::from_reader(tokenizer).context("while loading tokenizer")?;
+
+		let ccfg = ComputeConfig {
+			batch_size: 10,
+			eval_frac: 0.1,
+			eval_salt: "salt".into(),
+		};
+
+		let mcfg = GptModelConfig {
+			vocab_size: tokenizer.vocab_size(),
+			context_size: 256,
+			embed_dim: 768,
+			n_heads: 12,
+			head_dim: 64, // = 768 / 12
+			n_layers: 1,
+			embed_drop_rate: 0.1,
+			attention_drop_rate: 0.1,
+			shortcut_drop_rate: 0.1,
+		};
+
+		let mut model: GptModel<Autodiff<Cuda>> = mcfg.init(&device);
 
 		/*
-				let context = a;
+		let loader_train = DataLoaderBuilder::new(batcher.clone())
-				let desired = &b[b.len() - 1..];
+			.batch_size(ccfg.batch_size)
-				println!("{context:?} -> {desired:?}");
+			//.shuffle(config.seed)
+			.num_workers(5)
+			.build(Loader::new(&self.data_dir).context("while initializing loader")?);
 
-				let input = tokenizer.decode(context);
+		let loader_test = DataLoaderBuilder::new(batcher)
-				let target = tokenizer.decode(desired);
+			.batch_size(ccfg.batch_size)
-				println!("{input:?} -> {target:?}");
+			//.shuffle(config.seed)
+			.num_workers(5)
+			.build(Loader::new(&self.data_dir).context("while initializing loader")?);
+
+		let learner = LearnerBuilder::new("./tmp")
+			.metric_train_numeric(AccuracyMetric::new())
+			.metric_valid_numeric(AccuracyMetric::new())
+			.metric_train_numeric(LossMetric::new())
+			.metric_valid_numeric(LossMetric::new())
+			.with_file_checkpointer(CompactRecorder::new())
+			.learning_strategy(LearningStrategy::SingleDevice(device.clone()))
+			.num_epochs(10)
+			.summary()
+			.build(model, AdamConfig::new().init(), 1e-4);
+
+		learner.fit(loader_train, loader_test);
 		*/
 
-				if input_batch.len() >= batch_size {
+		// Initialize optimizer
-					let shape = [input_batch.len(), config.context_size];
+		let mut optim = AdamConfig::new().init();
+		let learning_rate = 1e-4;
 
-					let input = std::mem::replace(&mut input_batch, Vec::with_capacity(batch_size));
+		for epoch in 0..10 {
-					let input: Array2<u32> = Array2::from_shape_fn(shape, |(a, b)| input[a][b]);
+			debug!("Running epoch {epoch}");
 
-					#[expect(clippy::unwrap_used)]
+			// Training phase
-					let input: Tensor<Cuda, 2, Int> =
+			let mut train_loss_sum = 0.0;
-						Tensor::<_, 1, Int>::from_ints(input.as_slice().unwrap(), &device)
+			let mut train_total = 0;
-							.reshape(shape);
 
-					let output =
+			for batch in
-						std::mem::replace(&mut output_batch, Vec::with_capacity(batch_size));
+				TrainTestIterator::new(&self.data_dir, &ccfg, &mcfg, &tokenizer, false, &device)
-					let output: Array2<u32> = Array2::from_shape_fn(shape, |(a, b)| output[a][b]);
+					.context("while initializing reader")?
+			{
+				let batch = batch.context("while reading batch")?;
 
-					#[expect(clippy::unwrap_used)]
+				// Forward pass with gradients
-					let output: Tensor<Cuda, 2, Int> =
+				let output = model.forward_train(batch.inputs, batch.targets);
-						Tensor::<_, 1, Int>::from_ints(output.as_slice().unwrap(), &device)
-							.reshape(shape);
 
-					self.batch(&config, input, &model);
+				train_total += output.targets.dims()[0] as i32;
-				}
+				train_loss_sum += output.loss.clone().into_scalar().to_f32();
-			}
+
+				let grads = output.loss.backward();
+				let grads = GradientsParams::from_grads(grads, &model);
+
+				model = optim.step(learning_rate, model, grads);
 			}
 
-		if !input_batch.is_empty() {
+			let mut valid_loss_sum = 0.0;
-			let shape = [input_batch.len(), config.context_size];
+			let mut valid_total = 0;
 
-			let input = std::mem::replace(&mut input_batch, Vec::with_capacity(batch_size));
+			let mut n_eval = 0;
-			let input: Array2<u32> = Array2::from_shape_fn(shape, |(a, b)| input[a][b]);
+			debug!("Evaluating batches");
 
-			#[expect(clippy::unwrap_used)]
+			for batch in
-			let input: Tensor<Cuda, 2, Int> =
+				TrainTestIterator::new(&self.data_dir, &ccfg, &mcfg, &tokenizer, true, &device)
-				Tensor::<_, 1, Int>::from_ints(input.as_slice().unwrap(), &device).reshape(shape);
+					.context("while initializing reader")?
+			{
+				let batch = batch.context("while reading batch")?;
+				n_eval += batch.targets.shape()[0];
 
-			let output = std::mem::replace(&mut output_batch, Vec::with_capacity(batch_size));
+				// Forward pass without gradients
-			let output: Array2<u32> = Array2::from_shape_fn(shape, |(a, b)| output[a][b]);
+				let output = model.valid().forward_train(batch.inputs, batch.targets);
 
-			#[expect(clippy::unwrap_used)]
+				valid_total += output.targets.dims()[0] as i32;
-			let output: Tensor<Cuda, 2, Int> =
+				valid_loss_sum += output.loss.into_scalar().to_f32();
-				Tensor::<_, 1, Int>::from_ints(output.as_slice().unwrap(), &device).reshape(shape);
+			}
 
-			self.batch(&config, input, &model);
+			// Compute and log epoch results
+			let train_loss = if train_total > 0 {
+				train_loss_sum / train_total as f32
+			} else {
+				0.0
+			};
+			let valid_loss = if valid_total > 0 {
+				valid_loss_sum / valid_total as f32
+			} else {
+				0.0
+			};
+
+			info!(message = "Ran epoch", epoch, train_loss, valid_loss, n_eval);
 		}
 
 		Ok(())
 	}
+}
 
-	fn batch(&self, _cfg: &Config, input: Tensor<Cuda, 2, Int>, model: &GptModel<Cuda>) {
+//
-		let logits = model.forward(input);
+// MARK: model
-		println!("{:?}", logits.shape());
+//
-	}
-}
 
 /// Multihead attention.
 ///
@@ -315,7 +451,7 @@ impl<B: Backend> MultiheadAttention<B> {
 				},
 				device.clone(),
 				true,
-				[embedding_dim, total_dim].into(),
+				[total_dim, total_dim].into(),
 			),
 
 			dropout: Dropout { prob: dropout },
@@ -389,6 +525,7 @@ impl<B: Backend> MultiheadAttention<B> {
 		let mask = self
 			.utri_mask
 			.clone()
+			.slice([0..tokens, 0..tokens])
 			.unsqueeze_dim::<3>(0)
 			.unsqueeze_dim::<4>(0)
 			.expand(attn_scores.shape());
@@ -422,35 +559,50 @@ impl<B: Backend> MultiheadAttention<B> {
 	}
 }
 
-#[derive(Module, Debug)]
+#[derive(Config, Debug)]
-pub struct GptModel<B: Backend> {
+pub struct GptModelConfig {
-	embedder_tok: Embedding<B>,
+	/// Number of tokens
-	embedder_pos: Embedding<B>,
+	pub vocab_size: u32,
-	embedder_drop: Dropout,
 
-	trf_blocks: Vec<TransformerBlock<B>>,
+	/// Maximum number of input tokens with positional embeddings
-	final_norm: LayerNorm<B>,
+	pub context_size: usize,
-	out_head: Param<Tensor<B, 2>>,
+
+	/// Dimension of each token's embedding
+	pub embed_dim: usize,
+
+	/// Number of attention heads
+	pub n_heads: usize,
+
+	/// Dimension of each attn head
+	pub head_dim: usize,
+
+	/// Number of transformer blocks
+	pub n_layers: usize,
+
+	pub embed_drop_rate: f64,
+	pub attention_drop_rate: f64,
+	pub shortcut_drop_rate: f64,
 }
 
-impl<B: Backend> GptModel<B> {
+impl GptModelConfig {
-	pub fn new(cfg: &Config, device: &B::Device) -> Self {
+	pub fn init<B: Backend>(&self, device: &B::Device) -> GptModel<B> {
-		let out_head_shape = [cfg.embed_dim, cfg.vocab_size as usize];
+		let out_head_shape = [self.embed_dim, self.vocab_size as usize];
 
-		Self {
+		GptModel {
-			embedder_tok: EmbeddingConfig::new(cfg.vocab_size as usize, cfg.embed_dim).init(device),
+			embedder_tok: EmbeddingConfig::new(self.vocab_size as usize, self.embed_dim)
+				.init(device),
 
-			embedder_pos: EmbeddingConfig::new(cfg.context_size, cfg.embed_dim).init(device),
+			embedder_pos: EmbeddingConfig::new(self.context_size, self.embed_dim).init(device),
 
 			embedder_drop: Dropout {
-				prob: cfg.embed_drop_rate,
+				prob: self.embed_drop_rate,
 			},
 
-			trf_blocks: (0..cfg.n_layers)
+			trf_blocks: (0..self.n_layers)
-				.map(|_| TransformerBlock::new(cfg, device))
+				.map(|_| TransformerBlock::new(&self, device))
 				.collect(),
 
-			final_norm: LayerNormConfig::new(cfg.embed_dim).init(device),
+			final_norm: LayerNormConfig::new(self.embed_dim).init(device),
 
 			out_head: Param::uninitialized(
 				ParamId::new(),
@@ -464,13 +616,34 @@ impl<B: Backend> GptModel<B> {
 			),
 		}
 	}
+}
 
+#[derive(Debug, Clone)]
+pub struct TrainBatch<B: Backend> {
+	pub inputs: Tensor<B, 2, Int>,
+
+	/// Correct next token for each input
+	pub targets: Tensor<B, 1, Int>,
+}
+
+#[derive(Module, Debug)]
+pub struct GptModel<B: Backend> {
+	embedder_tok: Embedding<B>,
+	embedder_pos: Embedding<B>,
+	embedder_drop: Dropout,
+
+	trf_blocks: Vec<TransformerBlock<B>>,
+	final_norm: LayerNorm<B>,
+	out_head: Param<Tensor<B, 2>>,
+}
+
+impl<B: Backend> GptModel<B> {
 	pub fn forward(&self, input: Tensor<B, 2, Int>) -> Tensor<B, 3> {
 		let n_tokens = input.shape()[1];
 
 		let embed_tok = self.embedder_tok.forward(input.clone());
 		let embed_pos = self
-			.embedder_tok
+			.embedder_pos
 			.forward(Tensor::arange(0..n_tokens as i64, &input.device()).unsqueeze_dim(0));
 
 		let x = embed_tok + embed_pos;
@@ -481,6 +654,29 @@ impl<B: Backend> GptModel<B> {
 
 		return logits;
 	}
+
+	pub fn forward_train(
+		&self,
+		inputs: Tensor<B, 2, Int>,
+		targets: Tensor<B, 1, Int>,
+	) -> ClassificationOutput<B> {
+		// shape: [batch, n_tokens, n_vocabulary]
+		let output = self.forward(inputs);
+
+		// Get last token
+		// shape: [batch, n_vocabulary]
+		let output = output.slice_dim(1, -1).squeeze_dim::<2>(1);
+
+		let loss = CrossEntropyLossConfig::new()
+			.init(&targets.device())
+			.forward(output.clone(), targets.clone());
+
+		ClassificationOutput {
+			loss,
+			output,
+			targets,
+		}
+	}
 }
 
 #[derive(Module, Debug)]
@@ -498,7 +694,7 @@ pub struct TransformerBlock<B: Backend> {
 }
 
 impl<B: Backend> TransformerBlock<B> {
-	pub fn new(cfg: &Config, device: &B::Device) -> Self {
+	pub fn new(cfg: &GptModelConfig, device: &B::Device) -> Self {
 		Self {
 			attention: MultiheadAttention::new(
 				cfg.embed_dim,

crates/llmfs/src/data_reader.rs

@@ -25,10 +25,11 @@ pub enum DataReaderError {
 ///
 /// All parquet files have exactly one text column.
 /// No promises about this struct's behavior if this is not the case.
+#[derive(Clone)]
 pub struct DataReader {
-	rx: Receiver<Result<String, DataReaderError>>,
+	rx: Arc<Mutex<Receiver<Result<String, DataReaderError>>>>,
 	total_rows: usize,
-	consumed_rows: AtomicUsize,
+	consumed_rows: Arc<AtomicUsize>,
 }
 
 impl DataReader {
@@ -57,6 +58,15 @@ impl DataReader {
 					files.push(path);
 				}
 			}
+
+			files.sort_by_key(|a| {
+				a.file_name()
+					.map(|x| x.to_str())
+					.flatten()
+					.unwrap_or("")
+					.to_owned()
+			});
+
 			(Arc::new(Mutex::new(files)), total_rows)
 		};
 
@@ -147,9 +157,9 @@ impl DataReader {
 		});
 
 		Ok(Self {
-			rx,
+			rx: Arc::new(Mutex::new(rx)),
 			total_rows,
-			consumed_rows: AtomicUsize::new(0),
+			consumed_rows: Arc::new(AtomicUsize::new(0)),
 		})
 	}
 
@@ -157,7 +167,7 @@ impl DataReader {
 	/// Order is arbitrary.
 	/// Returns `None` when all rows have been read.
 	pub fn recv(&self) -> Option<Result<String, DataReaderError>> {
-		self.rx.recv().ok()
+		self.rx.lock().recv().ok()
 	}
 
 	//pub fn try_recv(&self) -> Result<Result<String, DataReaderError>, TryRecvError> {

crates/llmfs/src/logging.rs

@@ -12,8 +12,7 @@ use tracing_subscriber::{
 // MARK: loglevel
 //
 
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Deserialize, ValueEnum)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Deserialize, ValueEnum, Default)]
-#[derive(Default)]
 pub enum LogLevel {
 	Trace,
 	Debug,
@@ -23,7 +22,6 @@ pub enum LogLevel {
 	Error,
 }
 
-
 impl Display for LogLevel {
 	fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
 		match self {
@@ -71,7 +69,7 @@ impl From<LoggingConfig> for EnvFilter {
 				//
 				// Bins
 				//
-				format!("nanochat_rs={}", conf.nanochat),
+				format!("llmfs={}", conf.nanochat),
 				conf.other.to_string(),
 			]
 			.join(","),
@@ -216,8 +214,7 @@ pub enum LoggingTarget {
 }
 
 /// How to print logs
-#[derive(Debug, Clone, Copy, Deserialize)]
+#[derive(Debug, Clone, Copy, Deserialize, Default)]
-#[derive(Default)]
 pub enum LoggingFormat {
 	#[default]
 	Ansi,
@@ -225,7 +222,6 @@ pub enum LoggingFormat {
 	Json,
 }
 
-
 pub struct LoggingInitializer {
 	/// Log filter for printed logs
 	pub preset: LogFilterPreset,

crates/llmfs/src/main.rs

@@ -1,3 +1,5 @@
+#![recursion_limit = "256"]
+
 use clap::Parser;
 use indicatif::MultiProgress;
 use tracing::error;