red-candle 1.0.0.pre.1 → 1.0.0.pre.3

data/ext/candle/src/llm/mistral.rs

@@ -0,0 +1,325 @@
+use candle_core::{DType, Device, Result as CandleResult, Tensor};
+use candle_nn::VarBuilder;
+use candle_transformers::models::mistral::{Config, Model as MistralModel};
+use hf_hub::{api::tokio::Api, Repo};
+use tokenizers::Tokenizer;
+
+use super::{GenerationConfig, TextGeneration, TextGenerator, TokenizerWrapper};
+
+#[derive(Debug)]
+pub struct Mistral {
+    model: MistralModel,
+    tokenizer: TokenizerWrapper,
+    device: Device,
+    model_id: String,
+    eos_token_id: u32,
+}
+
+impl Mistral {
+    /// Clear the KV cache between generations
+    pub fn clear_kv_cache(&mut self) {
+        self.model.clear_kv_cache();
+    }
+    
+    /// Load a Mistral model from HuggingFace Hub
+    pub async fn from_pretrained(model_id: &str, device: Device) -> CandleResult<Self> {
+        let api = Api::new()
+            .map_err(|e| candle_core::Error::Msg(format!("Failed to create HF API: {}", e)))?;
+        
+        let repo = api.repo(Repo::model(model_id.to_string()));
+        
+        // Download model files
+        let config_filename = repo
+            .get("config.json")
+            .await
+            .map_err(|e| candle_core::Error::Msg(format!("Failed to download config: {}", e)))?;
+        
+        let tokenizer_filename = repo
+            .get("tokenizer.json")
+            .await
+            .map_err(|e| candle_core::Error::Msg(format!("Failed to download tokenizer: {}", e)))?;
+        
+        // Try different file patterns for model weights
+        let weights_filenames = if let Ok(single_file) = repo.get("model.safetensors").await {
+            vec![single_file]
+        } else if let Ok(consolidated_file) = repo.get("consolidated.safetensors").await {
+            // Some Mistral models use consolidated.safetensors
+            vec![consolidated_file]
+        } else {
+            // Try to find sharded model files
+            let mut sharded_files = Vec::new();
+            let mut index = 1;
+            loop {
+                // Try common shard counts
+                let mut found = false;
+                for total in [2, 3, 4, 5, 6, 7, 8] {
+                    let filename = format!("model-{:05}-of-{:05}.safetensors", index, total);
+                    if let Ok(file) = repo.get(&filename).await {
+                        sharded_files.push(file);
+                        found = true;
+                        break;
+                    }
+                }
+                if !found {
+                    break;
+                }
+                index += 1;
+            }
+            
+            if sharded_files.is_empty() {
+                // Try single pytorch_model.bin as last resort (though we prefer safetensors)
+                if let Ok(_pytorch_file) = repo.get("pytorch_model.bin").await {
+                    return Err(candle_core::Error::Msg(
+                        "Only safetensors format is supported. This model uses pytorch_model.bin format.".to_string()
+                    ));
+                } else {
+                    return Err(candle_core::Error::Msg(
+                        "Could not find model weights. Tried: model.safetensors, consolidated.safetensors, model-*-of-*.safetensors".to_string()
+                    ));
+                }
+            }
+            sharded_files
+        };
+        
+        // Load config
+        let config: Config = serde_json::from_reader(std::fs::File::open(config_filename)?)
+            .map_err(|e| candle_core::Error::Msg(format!("Failed to parse config: {}", e)))?;
+        
+        // Load tokenizer
+        let tokenizer = Tokenizer::from_file(tokenizer_filename)
+            .map_err(|e| candle_core::Error::Msg(format!("Failed to load tokenizer: {}", e)))?;
+        
+        let eos_token_id = tokenizer
+            .get_vocab(true)
+            .get("</s>")
+            .copied()
+            .unwrap_or(2);
+        
+        // Load model weights
+        let vb = unsafe {
+            VarBuilder::from_mmaped_safetensors(&weights_filenames, DType::F32, &device)?
+        };
+        
+        let model = MistralModel::new(&config, vb)?;
+        
+        Ok(Self {
+            model,
+            tokenizer: TokenizerWrapper::new(tokenizer),
+            device,
+            model_id: model_id.to_string(),
+            eos_token_id,
+        })
+    }
+
+    /// Create from existing components (useful for testing)
+    pub fn new(
+        model: MistralModel,
+        tokenizer: Tokenizer,
+        device: Device,
+        model_id: String,
+    ) -> Self {
+        let eos_token_id = tokenizer
+            .get_vocab(true)
+            .get("</s>")
+            .copied()
+            .unwrap_or(2);
+        
+        Self {
+            model,
+            tokenizer: TokenizerWrapper::new(tokenizer),
+            device,
+            model_id,
+            eos_token_id,
+        }
+    }
+
+    fn generate_tokens(
+        &mut self,
+        prompt_tokens: Vec<u32>,
+        config: &GenerationConfig,
+        mut callback: Option<impl FnMut(&str)>,
+    ) -> CandleResult<Vec<u32>> {
+        let mut text_gen = TextGeneration::from_config(config);
+        text_gen.set_eos_token_id(self.eos_token_id);
+        text_gen.set_tokens(prompt_tokens.clone());
+        
+        let mut all_tokens = prompt_tokens.clone();
+        let start_gen = all_tokens.len();
+        
+        for index in 0..config.max_length {
+            let context_size = if index > 0 { 1 } else { all_tokens.len() };
+            let start_pos = all_tokens.len().saturating_sub(context_size);
+            let ctxt = &all_tokens[start_pos..];
+            
+            let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
+            // Ensure input tensor is contiguous for Metal backend
+            let input = input.contiguous()?;
+            let logits = self.model.forward(&input, start_pos)?;
+            
+            // The model returns logits of shape [batch_size, seq_len, vocab_size]
+            // We need to get the logits for the last token only
+            let logits = logits.squeeze(0)?; // Remove batch dimension
+            let logits = if logits.dims().len() == 2 {
+                // If we still have [seq_len, vocab_size], take the last token
+                let seq_len = logits.dim(0)?;
+                logits.narrow(0, seq_len - 1, 1)?.squeeze(0)?
+            } else {
+                // Already [vocab_size]
+                logits
+            };
+            
+            // Convert to F32 for sampling if needed
+            let logits = logits.to_dtype(DType::F32)?;
+            
+            let next_token = text_gen.sample_next_token(
+                &logits,
+                Some((config.repetition_penalty, config.repetition_penalty_last_n)),
+            )?;
+            
+            all_tokens.push(next_token);
+            
+            // Stream callback
+            if let Some(ref mut cb) = callback {
+                let token_text = self.tokenizer.token_to_piece(next_token)?;
+                cb(&token_text);
+            }
+            
+            // Check stop conditions
+            if text_gen.should_stop(next_token, config.max_length) {
+                break;
+            }
+            
+            // Check stop sequences
+            let generated_text = self.tokenizer.decode(&all_tokens[start_gen..], true)?;
+            if text_gen.check_stop_sequences(&generated_text, &config.stop_sequences) {
+                break;
+            }
+        }
+        
+        Ok(if config.include_prompt {
+            all_tokens
+        } else {
+            all_tokens[start_gen..].to_vec()
+        })
+    }
+    
+    fn generate_tokens_decoded(
+        &mut self,
+        prompt_tokens: Vec<u32>,
+        config: &GenerationConfig,
+        mut callback: Option<impl FnMut(&str)>,
+    ) -> CandleResult<Vec<u32>> {
+        let mut text_gen = TextGeneration::from_config(config);
+        text_gen.set_eos_token_id(self.eos_token_id);
+        text_gen.set_tokens(prompt_tokens.clone());
+        
+        let mut all_tokens = prompt_tokens.clone();
+        let start_gen = all_tokens.len();
+        
+        // For incremental decoding
+        let mut previously_decoded = String::new();
+        
+        for index in 0..config.max_length {
+            let context_size = if index > 0 { 1 } else { all_tokens.len() };
+            let start_pos = all_tokens.len().saturating_sub(context_size);
+            let ctxt = &all_tokens[start_pos..];
+            
+            let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
+            // Ensure input tensor is contiguous for Metal backend
+            let input = input.contiguous()?;
+            let logits = self.model.forward(&input, start_pos)?;
+            
+            // The model returns logits of shape [batch_size, seq_len, vocab_size]
+            // We need to get the logits for the last token only
+            let logits = logits.squeeze(0)?; // Remove batch dimension
+            let logits = if logits.dims().len() == 2 {
+                // If we still have [seq_len, vocab_size], take the last token
+                let seq_len = logits.dim(0)?;
+                logits.narrow(0, seq_len - 1, 1)?.squeeze(0)?
+            } else {
+                // Already [vocab_size]
+                logits
+            };
+            
+            // Convert to F32 for sampling if needed
+            let logits = logits.to_dtype(DType::F32)?;
+            
+            let next_token = text_gen.sample_next_token(
+                &logits,
+                Some((config.repetition_penalty, config.repetition_penalty_last_n)),
+            )?;
+            
+            all_tokens.push(next_token);
+            
+            // Stream callback with incremental decoding
+            if let Some(ref mut cb) = callback {
+                // Decode all generated tokens so far
+                let current_decoded = self.tokenizer.decode(&all_tokens[start_gen..], true)?;
+                
+                // Only emit the new text since last callback
+                if current_decoded.len() > previously_decoded.len() {
+                    let new_text = &current_decoded[previously_decoded.len()..];
+                    cb(new_text);
+                    previously_decoded = current_decoded;
+                }
+            }
+            
+            // Check stop conditions
+            if text_gen.should_stop(next_token, config.max_length) {
+                break;
+            }
+            
+            // Check stop sequences
+            let generated_text = if callback.is_some() {
+                previously_decoded.clone()
+            } else {
+                self.tokenizer.decode(&all_tokens[start_gen..], true)?
+            };
+            
+            if text_gen.check_stop_sequences(&generated_text, &config.stop_sequences) {
+                break;
+            }
+        }
+        
+        Ok(if config.include_prompt {
+            all_tokens
+        } else {
+            all_tokens[start_gen..].to_vec()
+        })
+    }
+}
+
+impl TextGenerator for Mistral {
+    fn generate(
+        &mut self,
+        prompt: &str,
+        config: &GenerationConfig,
+    ) -> CandleResult<String> {
+        let prompt_tokens = self.tokenizer.encode(prompt, true)?;
+        let output_tokens = self.generate_tokens(prompt_tokens, config, None::<fn(&str)>)?;
+        self.tokenizer.decode(&output_tokens, true)
+    }
+
+    fn generate_stream(
+        &mut self,
+        prompt: &str,
+        config: &GenerationConfig,
+        mut callback: impl FnMut(&str),
+    ) -> CandleResult<String> {
+        let prompt_tokens = self.tokenizer.encode(prompt, true)?;
+        let output_tokens = self.generate_tokens_decoded(prompt_tokens, config, Some(&mut callback))?;
+        self.tokenizer.decode(&output_tokens, true)
+    }
+
+    fn model_name(&self) -> &str {
+        &self.model_id
+    }
+
+    fn device(&self) -> &Device {
+        &self.device
+    }
+    
+    fn clear_cache(&mut self) {
+        self.clear_kv_cache();
+    }
+}

data/ext/candle/src/llm/mod.rs

@@ -0,0 +1,68 @@
+use candle_core::{Device, Result as CandleResult};
+use tokenizers::Tokenizer;
+
+pub mod mistral;
+pub mod generation_config;
+pub mod text_generation;
+
+pub use generation_config::GenerationConfig;
+pub use text_generation::TextGeneration;
+
+/// Trait for text generation models
+pub trait TextGenerator: Send + Sync {
+    /// Generate text from a prompt
+    fn generate(
+        &mut self,
+        prompt: &str,
+        config: &GenerationConfig,
+    ) -> CandleResult<String>;
+
+    /// Generate text with streaming callback
+    fn generate_stream(
+        &mut self,
+        prompt: &str,
+        config: &GenerationConfig,
+        callback: impl FnMut(&str),
+    ) -> CandleResult<String>;
+
+    /// Get the model's name
+    fn model_name(&self) -> &str;
+
+    /// Get the device the model is running on
+    fn device(&self) -> &Device;
+    
+    /// Clear any cached state (like KV cache)
+    fn clear_cache(&mut self);
+}
+
+/// Common structure for managing tokenizer
+#[derive(Debug)]
+pub struct TokenizerWrapper {
+    tokenizer: Tokenizer,
+}
+
+impl TokenizerWrapper {
+    pub fn new(tokenizer: Tokenizer) -> Self {
+        Self { tokenizer }
+    }
+
+    pub fn encode(&self, text: &str, add_special_tokens: bool) -> CandleResult<Vec<u32>> {
+        let encoding = self.tokenizer
+            .encode(text, add_special_tokens)
+            .map_err(|e| candle_core::Error::Msg(format!("Tokenizer error: {}", e)))?;
+        Ok(encoding.get_ids().to_vec())
+    }
+
+    pub fn decode(&self, tokens: &[u32], skip_special_tokens: bool) -> CandleResult<String> {
+        self.tokenizer
+            .decode(tokens, skip_special_tokens)
+            .map_err(|e| candle_core::Error::Msg(format!("Tokenizer decode error: {}", e)))
+    }
+
+    pub fn token_to_piece(&self, token: u32) -> CandleResult<String> {
+        self.tokenizer
+            .id_to_token(token)
+            .map(|s| s.to_string())
+            .ok_or_else(|| candle_core::Error::Msg(format!("Unknown token id: {}", token)))
+    }
+}

data/ext/candle/src/llm/text_generation.rs

@@ -0,0 +1,141 @@
+use candle_core::{Result as CandleResult, Tensor};
+use candle_transformers::generation::LogitsProcessor;
+use rand::{rngs::StdRng, SeedableRng};
+
+use super::GenerationConfig;
+
+/// Helper struct for text generation process
+pub struct TextGeneration {
+    #[allow(dead_code)]
+    rng: StdRng,
+    logits_processor: LogitsProcessor,
+    tokens: Vec<u32>,
+    eos_token_id: Option<u32>,
+}
+
+impl TextGeneration {
+    pub fn new(
+        seed: u64,
+        temperature: Option<f64>,
+        top_p: Option<f64>,
+        _top_k: Option<usize>,
+        _repetition_penalty: f32,
+        _repetition_penalty_last_n: usize,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temperature, top_p);
+        
+        Self {
+            rng: StdRng::seed_from_u64(seed),
+            logits_processor,
+            tokens: Vec::new(),
+            eos_token_id: None,
+        }
+    }
+
+    pub fn from_config(config: &GenerationConfig) -> Self {
+        Self::new(
+            config.seed,
+            Some(config.temperature),
+            config.top_p,
+            config.top_k,
+            config.repetition_penalty,
+            config.repetition_penalty_last_n,
+        )
+    }
+
+    pub fn set_eos_token_id(&mut self, eos_token_id: u32) {
+        self.eos_token_id = Some(eos_token_id);
+    }
+
+    pub fn set_tokens(&mut self, tokens: Vec<u32>) {
+        self.tokens = tokens;
+    }
+
+    pub fn get_tokens(&self) -> &[u32] {
+        &self.tokens
+    }
+
+    pub fn push_token(&mut self, token: u32) {
+        self.tokens.push(token);
+    }
+
+    /// Apply repetition penalty to logits
+    pub fn apply_repetition_penalty(
+        &self,
+        logits: &mut Tensor,
+        penalty: f32,
+        context_size: usize,
+    ) -> CandleResult<()> {
+        if penalty == 1.0 {
+            return Ok(());
+        }
+
+        let device = logits.device();
+        let vocab_size = logits.dims1()?;
+        
+        // Get the context tokens to apply penalty to
+        let start = self.tokens.len().saturating_sub(context_size);
+        let context_tokens = &self.tokens[start..];
+        
+        // Apply penalty to tokens that appear in the context
+        let mut logits_vec = logits.to_vec1::<f32>()?;
+        for &token in context_tokens {
+            if (token as usize) < vocab_size {
+                let idx = token as usize;
+                if logits_vec[idx] > 0.0 {
+                    logits_vec[idx] /= penalty;
+                } else {
+                    logits_vec[idx] *= penalty;
+                }
+            }
+        }
+        
+        *logits = Tensor::from_vec(logits_vec, vocab_size, device)?;
+        Ok(())
+    }
+
+    /// Sample next token from logits
+    pub fn sample_next_token(
+        &mut self,
+        logits: &Tensor,
+        repetition_penalty: Option<(f32, usize)>,
+    ) -> CandleResult<u32> {
+        let mut logits = logits.clone();
+        
+        // Apply repetition penalty if specified
+        if let Some((penalty, last_n)) = repetition_penalty {
+            self.apply_repetition_penalty(&mut logits, penalty, last_n)?;
+        }
+        
+        // Sample token
+        let next_token = self.logits_processor.sample(&logits)?;
+        self.tokens.push(next_token);
+        
+        Ok(next_token)
+    }
+
+    /// Check if we should stop generation
+    pub fn should_stop(&self, token: u32, max_length: usize) -> bool {
+        if self.tokens.len() >= max_length {
+            return true;
+        }
+        
+        if let Some(eos) = self.eos_token_id {
+            if token == eos {
+                return true;
+            }
+        }
+        
+        false
+    }
+
+    /// Check if the generated text ends with any stop sequence
+    pub fn check_stop_sequences(&self, text: &str, stop_sequences: &[String]) -> bool {
+        for seq in stop_sequences {
+            if text.ends_with(seq) {
+                return true;
+            }
+        }
+        false
+    }
+}