|
| 1 | +// An fp8 implementation of llama using accelerated |
| 2 | +// fp8 matmuls in cublaslt |
| 3 | + |
| 4 | +use anyhow::{bail, Error as E, Result}; |
| 5 | +use clap::{Parser, ValueEnum}; |
| 6 | + |
| 7 | +use candle::{DType, Tensor}; |
| 8 | +use candle_nn::VarBuilder; |
| 9 | +use candle_transformers::generation::{LogitsProcessor, Sampling}; |
| 10 | +use hf_hub::{api::sync::Api, Repo, RepoType}; |
| 11 | +use std::io::Write; |
| 12 | + |
| 13 | +use candle_transformers::models::llama_fp8 as model; |
| 14 | +use model::{Llama, LlamaConfig}; |
| 15 | + |
| 16 | +const EOS_TOKEN: &str = "</s>"; |
| 17 | +const DEFAULT_PROMPT: &str = "My favorite theorem is "; |
| 18 | + |
| 19 | +#[derive(Clone, Debug, Copy, PartialEq, Eq, ValueEnum)] |
| 20 | +enum Which { |
| 21 | + #[value(name = "RedHat-3.1-8B-Instruct-FP8")] |
| 22 | + #[allow(non_camel_case_types)] |
| 23 | + RedHat31_8b_Instruct_FP8, |
| 24 | +} |
| 25 | + |
| 26 | +#[derive(Parser, Debug)] |
| 27 | +#[command(author, version, about, long_about = None)] |
| 28 | +struct Args { |
| 29 | + /// Run on CPU rather than on GPU. |
| 30 | + #[arg(long)] |
| 31 | + cpu: bool, |
| 32 | + |
| 33 | + /// The temperature used to generate samples. |
| 34 | + #[arg(long, default_value_t = 0.8)] |
| 35 | + temperature: f64, |
| 36 | + |
| 37 | + /// Nucleus sampling probability cutoff. |
| 38 | + #[arg(long)] |
| 39 | + top_p: Option<f64>, |
| 40 | + |
| 41 | + /// Only sample among the top K samples. |
| 42 | + #[arg(long)] |
| 43 | + top_k: Option<usize>, |
| 44 | + |
| 45 | + /// The seed to use when generating random samples. |
| 46 | + #[arg(long, default_value_t = 299792458)] |
| 47 | + seed: u64, |
| 48 | + |
| 49 | + /// The length of the sample to generate (in tokens). |
| 50 | + #[arg(short = 'n', long, default_value_t = 10000)] |
| 51 | + sample_len: usize, |
| 52 | + |
| 53 | + /// Disable the key-value cache. |
| 54 | + #[arg(long)] |
| 55 | + no_kv_cache: bool, |
| 56 | + |
| 57 | + /// The initial prompt. |
| 58 | + #[arg(long)] |
| 59 | + prompt: Option<String>, |
| 60 | + |
| 61 | + /// Use different dtype than f16 |
| 62 | + #[arg(long)] |
| 63 | + dtype: Option<String>, |
| 64 | + |
| 65 | + /// Enable tracing (generates a trace-timestamp.json file). |
| 66 | + #[arg(long)] |
| 67 | + tracing: bool, |
| 68 | + |
| 69 | + #[arg(long)] |
| 70 | + model_id: Option<String>, |
| 71 | + |
| 72 | + #[arg(long)] |
| 73 | + revision: Option<String>, |
| 74 | + |
| 75 | + /// The model size to use. |
| 76 | + #[arg(long, default_value = "v3")] |
| 77 | + which: Which, |
| 78 | + |
| 79 | + #[arg(long)] |
| 80 | + use_flash_attn: bool, |
| 81 | + |
| 82 | + /// Penalty to be applied for repeating tokens, 1. means no penalty. |
| 83 | + #[arg(long, default_value_t = 1.1)] |
| 84 | + repeat_penalty: f32, |
| 85 | + |
| 86 | + /// The context size to consider for the repeat penalty. |
| 87 | + #[arg(long, default_value_t = 128)] |
| 88 | + repeat_last_n: usize, |
| 89 | +} |
| 90 | + |
| 91 | +fn main() -> anyhow::Result<()> { |
| 92 | + use tokenizers::Tokenizer; |
| 93 | + use tracing_chrome::ChromeLayerBuilder; |
| 94 | + use tracing_subscriber::prelude::*; |
| 95 | + |
| 96 | + let args = Args::parse(); |
| 97 | + let _guard = if args.tracing { |
| 98 | + let (chrome_layer, guard) = ChromeLayerBuilder::new().build(); |
| 99 | + tracing_subscriber::registry().with(chrome_layer).init(); |
| 100 | + Some(guard) |
| 101 | + } else { |
| 102 | + None |
| 103 | + }; |
| 104 | + |
| 105 | + let device = candle_examples::device(args.cpu)?; |
| 106 | + let dtype = match args.dtype.as_deref() { |
| 107 | + Some("f16") => DType::F16, |
| 108 | + Some("bf16") => DType::BF16, |
| 109 | + Some("f32") => DType::F32, |
| 110 | + Some(dtype) => bail!("Unsupported dtype {dtype}"), |
| 111 | + None => DType::F16, |
| 112 | + }; |
| 113 | + |
| 114 | + let (llama, tokenizer_filename, mut cache, config) = { |
| 115 | + let api = Api::new()?; |
| 116 | + let model_id = args.model_id.unwrap_or_else(|| { |
| 117 | + let str = match args.which { |
| 118 | + Which::RedHat31_8b_Instruct_FP8 => "RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8", |
| 119 | + }; |
| 120 | + |
| 121 | + str.to_string() |
| 122 | + }); |
| 123 | + |
| 124 | + println!("loading the model weights from {model_id}"); |
| 125 | + let revision = args.revision.unwrap_or("main".to_string()); |
| 126 | + let api = api.repo(Repo::with_revision(model_id, RepoType::Model, revision)); |
| 127 | + |
| 128 | + let tokenizer_filename = api.get("tokenizer.json")?; |
| 129 | + let config_filename = api.get("config.json")?; |
| 130 | + let config: LlamaConfig = serde_json::from_slice(&std::fs::read(config_filename)?)?; |
| 131 | + let config = config.into_config(args.use_flash_attn); |
| 132 | + |
| 133 | + let filenames = match args.which { |
| 134 | + _ => candle_examples::hub_load_safetensors(&api, "model.safetensors.index.json")?, |
| 135 | + }; |
| 136 | + |
| 137 | + let cache = model::Cache::new(!args.no_kv_cache, dtype, &config, &device)?; |
| 138 | + |
| 139 | + let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? }; |
| 140 | + (Llama::load(vb, &config)?, tokenizer_filename, cache, config) |
| 141 | + }; |
| 142 | + let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?; |
| 143 | + let eos_token_id = config.eos_token_id.or_else(|| { |
| 144 | + tokenizer |
| 145 | + .token_to_id(EOS_TOKEN) |
| 146 | + .map(model::LlamaEosToks::Single) |
| 147 | + }); |
| 148 | + let prompt = args.prompt.as_ref().map_or(DEFAULT_PROMPT, |p| p.as_str()); |
| 149 | + let mut tokens = tokenizer |
| 150 | + .encode(prompt, true) |
| 151 | + .map_err(E::msg)? |
| 152 | + .get_ids() |
| 153 | + .to_vec(); |
| 154 | + let mut tokenizer = candle_examples::token_output_stream::TokenOutputStream::new(tokenizer); |
| 155 | + |
| 156 | + println!("starting the inference loop"); |
| 157 | + print!("{prompt}"); |
| 158 | + let mut logits_processor = { |
| 159 | + let temperature = args.temperature; |
| 160 | + let sampling = if temperature <= 0. { |
| 161 | + Sampling::ArgMax |
| 162 | + } else { |
| 163 | + match (args.top_k, args.top_p) { |
| 164 | + (None, None) => Sampling::All { temperature }, |
| 165 | + (Some(k), None) => Sampling::TopK { k, temperature }, |
| 166 | + (None, Some(p)) => Sampling::TopP { p, temperature }, |
| 167 | + (Some(k), Some(p)) => Sampling::TopKThenTopP { k, p, temperature }, |
| 168 | + } |
| 169 | + }; |
| 170 | + LogitsProcessor::from_sampling(args.seed, sampling) |
| 171 | + }; |
| 172 | + |
| 173 | + let mut start_gen = std::time::Instant::now(); |
| 174 | + let mut index_pos = 0; |
| 175 | + let mut token_generated = 0; |
| 176 | + for index in 0..args.sample_len { |
| 177 | + let (context_size, context_index) = if cache.use_kv_cache && index > 0 { |
| 178 | + (1, index_pos) |
| 179 | + } else { |
| 180 | + (tokens.len(), 0) |
| 181 | + }; |
| 182 | + if index == 1 { |
| 183 | + start_gen = std::time::Instant::now() |
| 184 | + } |
| 185 | + let ctxt = &tokens[tokens.len().saturating_sub(context_size)..]; |
| 186 | + let input = Tensor::new(ctxt, &device)?.unsqueeze(0)?; |
| 187 | + let logits = llama.forward(&input, context_index, &mut cache)?; |
| 188 | + let logits = logits.squeeze(0)?; |
| 189 | + let logits = if args.repeat_penalty == 1. { |
| 190 | + logits |
| 191 | + } else { |
| 192 | + let start_at = tokens.len().saturating_sub(args.repeat_last_n); |
| 193 | + candle_transformers::utils::apply_repeat_penalty( |
| 194 | + &logits, |
| 195 | + args.repeat_penalty, |
| 196 | + &tokens[start_at..], |
| 197 | + )? |
| 198 | + }; |
| 199 | + index_pos += ctxt.len(); |
| 200 | + |
| 201 | + let next_token = logits_processor.sample(&logits)?; |
| 202 | + token_generated += 1; |
| 203 | + tokens.push(next_token); |
| 204 | + |
| 205 | + match eos_token_id { |
| 206 | + Some(model::LlamaEosToks::Single(eos_tok_id)) if next_token == eos_tok_id => { |
| 207 | + break; |
| 208 | + } |
| 209 | + Some(model::LlamaEosToks::Multiple(ref eos_ids)) if eos_ids.contains(&next_token) => { |
| 210 | + break; |
| 211 | + } |
| 212 | + _ => (), |
| 213 | + } |
| 214 | + if let Some(t) = tokenizer.next_token(next_token)? { |
| 215 | + print!("{t}"); |
| 216 | + std::io::stdout().flush()?; |
| 217 | + } |
| 218 | + } |
| 219 | + if let Some(rest) = tokenizer.decode_rest().map_err(E::msg)? { |
| 220 | + print!("{rest}"); |
| 221 | + } |
| 222 | + let dt = start_gen.elapsed(); |
| 223 | + println!( |
| 224 | + "\n\n{} tokens generated ({} token/s)\n", |
| 225 | + token_generated, |
| 226 | + (token_generated - 1) as f64 / dt.as_secs_f64(), |
| 227 | + ); |
| 228 | + Ok(()) |
| 229 | +} |
0 commit comments