red-candle 1.8.0-aarch64-linux

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

@@ -0,0 +1,309 @@
+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 {
+    pub fn eos_token_id(&self) -> u32 {
+        self.eos_token_id
+    }
+
+    /// Clear the KV cache between generations
+    pub fn clear_kv_cache(&mut self) {
+        self.model.clear_kv_cache();
+    }
+    
+    /// Get the tokenizer
+    pub fn tokenizer(&self) -> &TokenizerWrapper {
+        &self.tokenizer
+    }
+    
+    /// Load a Mistral model from HuggingFace Hub with optional custom tokenizer
+    pub async fn from_pretrained_with_tokenizer(model_id: &str, device: Device, tokenizer_source: Option<&str>) -> 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)))?;
+        
+        // Download tokenizer from custom source if provided, otherwise from model repo
+        let tokenizer = if let Some(tokenizer_id) = tokenizer_source {
+            let tokenizer_repo = api.repo(Repo::model(tokenizer_id.to_string()));
+            let tokenizer_filename = tokenizer_repo
+                .get("tokenizer.json")
+                .await
+                .map_err(|e| {
+                    let error_msg = if e.to_string().contains("404") || e.to_string().contains("Not Found") {
+                        format!("Tokenizer file 'tokenizer.json' not found in repository '{}'. The repository may not have a tokenizer.json file or may use a different format (e.g., tokenizer.model for SentencePiece).", tokenizer_id)
+                    } else if e.to_string().contains("401") || e.to_string().contains("Unauthorized") {
+                        format!("Authentication required to access tokenizer '{}'. You may need to set HF_TOKEN environment variable with a valid Hugging Face token.", tokenizer_id)
+                    } else if e.to_string().contains("timed out") || e.to_string().contains("connection") {
+                        format!("Network error downloading tokenizer from '{}': {}. Please check your internet connection.", tokenizer_id, e)
+                    } else {
+                        format!("Failed to download tokenizer from '{}': {}", tokenizer_id, e)
+                    };
+                    candle_core::Error::Msg(error_msg)
+                })?;
+            Tokenizer::from_file(tokenizer_filename)
+                .map_err(|e| candle_core::Error::Msg(format!("Failed to load tokenizer file: {}", e)))?
+        } else {
+            let tokenizer_filename = repo
+                .get("tokenizer.json")
+                .await
+                .map_err(|e| {
+                    let error_msg = if e.to_string().contains("404") || e.to_string().contains("Not Found") {
+                        format!("No tokenizer found in model repository '{}'. The model may not include a tokenizer. Try specifying a tokenizer explicitly using the 'tokenizer' parameter, e.g.: from_pretrained('{}', tokenizer: 'mistralai/Mistral-7B-Instruct-v0.2')", model_id, model_id)
+                    } else if e.to_string().contains("401") || e.to_string().contains("Unauthorized") {
+                        format!("Authentication required to access model '{}'. You may need to set HF_TOKEN environment variable with a valid Hugging Face token.", model_id)
+                    } else if e.to_string().contains("timed out") || e.to_string().contains("connection") {
+                        format!("Network error downloading tokenizer: {}. Please check your internet connection.", e)
+                    } else {
+                        format!("Failed to download tokenizer: {}", e)
+                    };
+                    candle_core::Error::Msg(error_msg)
+                })?;
+            Tokenizer::from_file(tokenizer_filename)
+                .map_err(|e| candle_core::Error::Msg(format!("Failed to load tokenizer file: {}", 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
+            // NOTE: This uses a brute-force approach, trying common shard counts.
+            // A better approach would be to read model.safetensors.index.json which
+            // contains the exact file list, but this works for most models (≤8 shards).
+            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)))?;
+        
+        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,
+        })
+    }
+
+    /// Load a Mistral model from HuggingFace Hub (backwards compatibility)
+    pub async fn from_pretrained(model_id: &str, device: Device) -> CandleResult<Self> {
+        Self::from_pretrained_with_tokenizer(model_id, device, None).await
+    }
+
+    /// 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::new(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)?;
+            
+            all_tokens.push(next_token);
+            
+            // Stream callback
+            if let Some(ref mut cb) = callback {
+                if config.debug_tokens {
+                    // In debug mode, only show debug tokens
+                    let token_piece = self.tokenizer.token_to_piece(next_token)?;
+                    cb(&format!("[{}:{}]", next_token, token_piece));
+                } else {
+                    // Normal mode: use incremental decoding for proper text
+                    let decoded_text = self.tokenizer.decode_incremental(&all_tokens, all_tokens.len() - 1)?;
+                    cb(&decoded_text);
+                }
+            }
+            
+            // Check stop conditions
+            if text_gen.should_stop(next_token, config.max_length) {
+                break;
+            }
+            
+            // Check if constraint is satisfied (early stopping)
+            if config.stop_on_constraint_satisfaction {
+                let satisfied = if config.stop_on_match {
+                    text_gen.is_constraint_satisfied_stop_on_match()
+                } else {
+                    text_gen.is_constraint_satisfied()
+                };
+                if satisfied {
+                    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()
+        })
+    }
+}
+
+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)>)?;
+        
+        if config.debug_tokens {
+            self.tokenizer.format_tokens_with_debug(&output_tokens)
+        } else {
+            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(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 +1,49 @@
+use candle_core::{Device, Result as CandleResult};
+
+pub mod mistral;
+pub mod llama;
+pub mod gemma;
+pub mod qwen;
+pub mod qwen3;
+pub mod phi;
+pub mod granite;
+pub mod granitemoehybrid;
+pub mod glm4;
+pub mod generation_config;
+pub mod text_generation;
+pub mod quantized_gguf;
+
+pub use generation_config::GenerationConfig;
+pub use text_generation::TextGeneration;
+pub use quantized_gguf::QuantizedGGUF;
+pub use crate::tokenizer::TokenizerWrapper;
+
+#[cfg(test)]
+mod constrained_generation_test;
+
+/// 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);
+}