|
| 1 | +use crate::{Operators, RandomSample, Weights}; |
| 2 | +use common::Distribution; |
| 3 | +use gguf::GGufModel; |
| 4 | +use minicpm3::{ext::ggml_quants::f16, MiniCPM3Request, MiniCPM3Storage, Minicpm3Worker, Tensor}; |
| 5 | +use operators::{ |
| 6 | + all_reduce::common_cpu::Operator as AllReduce, |
| 7 | + common_cpu::{InprocNode, ThisThread}, |
| 8 | + random_sample::{KVPair, SampleArgs}, |
| 9 | + Blob, |
| 10 | +}; |
| 11 | +use regex::Regex; |
| 12 | +use std::{ |
| 13 | + iter::zip, |
| 14 | + ptr::copy_nonoverlapping, |
| 15 | + slice::from_raw_parts_mut, |
| 16 | + sync::{Arc, Barrier}, |
| 17 | + thread, |
| 18 | +}; |
| 19 | +use test_utils::{test_infer_paralle, Inference, Task, TokenizerAndPrompt, WorkerSeed}; |
| 20 | + |
| 21 | +type Worker<'w> = Minicpm3Worker<Operators<InprocNode<usize>, AllReduce>, Weights<'w>>; |
| 22 | + |
| 23 | +#[test] |
| 24 | +fn test_infer() { |
| 25 | + std::env::set_var( |
| 26 | + "TEST_MODEL", |
| 27 | + "/home/ztf/cpm/Origin-MiniCPM3-4B-v0.0-F16.gguf", |
| 28 | + ); |
| 29 | + let Some(Inference { |
| 30 | + model, |
| 31 | + devices, |
| 32 | + mut prompt, |
| 33 | + as_user, |
| 34 | + temperature, |
| 35 | + top_p, |
| 36 | + top_k, |
| 37 | + max_steps, |
| 38 | + }) = Inference::load() |
| 39 | + else { |
| 40 | + return; |
| 41 | + }; |
| 42 | + prompt = "我".to_owned(); |
| 43 | + let gguf = GGufModel::read(model.iter().map(|s| &**s)); |
| 44 | + |
| 45 | + let TokenizerAndPrompt { |
| 46 | + eos, |
| 47 | + tokenizer, |
| 48 | + prompt, |
| 49 | + } = TokenizerAndPrompt::new(&gguf, prompt, as_user); |
| 50 | + |
| 51 | + let model = MiniCPM3Storage::from_gguf(&gguf); |
| 52 | + println!("{:?}", model.meta); |
| 53 | + |
| 54 | + let sample_args = SampleArgs::new(temperature, top_p, top_k).expect("invalid sample args"); |
| 55 | + println!("{sample_args:?}"); |
| 56 | + |
| 57 | + let lens = devices |
| 58 | + .map(|devices| { |
| 59 | + Regex::new(r"\d+") |
| 60 | + .unwrap() |
| 61 | + .find_iter(&devices) |
| 62 | + .map(|c| c.as_str().parse().unwrap()) |
| 63 | + .collect() |
| 64 | + }) |
| 65 | + .unwrap_or_else(|| vec![1]); |
| 66 | + let dist = lens.iter().sum(); |
| 67 | + println!("distribution: {lens:?}"); |
| 68 | + |
| 69 | + let (seeds, senders) = WorkerSeed::new(InprocNode::new(lens.len())); |
| 70 | + let barrier = Arc::new(Barrier::new(dist + 1)); |
| 71 | + thread::scope(|s| { |
| 72 | + let _workers = zip(lens, seeds) |
| 73 | + .enumerate() |
| 74 | + .scan(0, |start, (id, (len, seed))| { |
| 75 | + let dist = Distribution::new(*start, len, dist); |
| 76 | + *start += len; |
| 77 | + |
| 78 | + let meta = model.meta.distribute(dist); |
| 79 | + let model = &model; |
| 80 | + let barrier = barrier.clone(); |
| 81 | + Some(s.spawn(move || { |
| 82 | + let WorkerSeed { node, tasks } = seed; |
| 83 | + let weights = Weights::new(model, dist); |
| 84 | + let mut worker = Worker::new(id, &node, meta.clone(), weights); |
| 85 | + let mut cache = meta.kv_cache(meta.nctx).map(Blob::new); |
| 86 | + let sin_cos = <Operators as minicpm3::Operators>::build_sin_cos( |
| 87 | + meta.dt_embd, |
| 88 | + meta.nctx, |
| 89 | + meta.dh, |
| 90 | + &ThisThread, |
| 91 | + ); |
| 92 | + |
| 93 | + let sample = RandomSample::new(&node); |
| 94 | + let indices = RandomSample::build_indices(model.meta.nvoc, &ThisThread); |
| 95 | + let mut pair = KVPair::new(0, f16::ZERO); |
| 96 | + let mut pairs = Tensor::kv_pair_vec(1, |_| unsafe { |
| 97 | + from_raw_parts_mut(&mut pair as *mut _ as *mut u8, size_of_val(&pair)) |
| 98 | + }); |
| 99 | + |
| 100 | + barrier.wait(); |
| 101 | + for task in tasks { |
| 102 | + let Task { |
| 103 | + nt, |
| 104 | + pos, |
| 105 | + embd, |
| 106 | + next, |
| 107 | + } = task; |
| 108 | + let mut embd = meta.embd(nt).map(|size| { |
| 109 | + let mut blob = Blob::new(size); |
| 110 | + unsafe { copy_nonoverlapping(embd, blob.as_mut_ptr(), size) }; |
| 111 | + blob |
| 112 | + }); |
| 113 | + let mut logits = meta.logits(if id == 0 { 1 } else { 0 }).map(Blob::new); |
| 114 | + worker |
| 115 | + .launch( |
| 116 | + minicpm3::MiniCPM3Args { |
| 117 | + embd: embd.map_slice_mut(), |
| 118 | + logits: logits.map_slice_mut(), |
| 119 | + sin_cos: sin_cos.map_slice(), |
| 120 | + requests: vec![MiniCPM3Request { |
| 121 | + cache: cache.map_slice_mut(), |
| 122 | + seq_len: nt, |
| 123 | + out_len: if id == 0 { 1 } else { 0 }, |
| 124 | + pos, |
| 125 | + }], |
| 126 | + num_tokens: nt, |
| 127 | + max_seq_len: nt, |
| 128 | + max_att_len: nt + pos, |
| 129 | + }, |
| 130 | + &mut [], |
| 131 | + &ThisThread, |
| 132 | + ) |
| 133 | + .unwrap(); |
| 134 | + if id == 0 { |
| 135 | + sample |
| 136 | + .launch( |
| 137 | + &mut pairs, |
| 138 | + &logits, |
| 139 | + &indices, |
| 140 | + sample_args, |
| 141 | + &mut [], |
| 142 | + &ThisThread, |
| 143 | + ) |
| 144 | + .unwrap(); |
| 145 | + next.send(pair.idx() as _).unwrap() |
| 146 | + } |
| 147 | + } |
| 148 | + })) |
| 149 | + }) |
| 150 | + .collect::<Vec<_>>(); |
| 151 | + |
| 152 | + let senders = senders.into_boxed_slice(); |
| 153 | + barrier.wait(); |
| 154 | + test_infer_paralle( |
| 155 | + senders, |
| 156 | + test_utils::AboutToken { |
| 157 | + tokenizer, |
| 158 | + token_embd: model.token_embd, |
| 159 | + nvoc: model.meta.nvoc, |
| 160 | + eos, |
| 161 | + }, |
| 162 | + &prompt, |
| 163 | + max_steps, |
| 164 | + ) |
| 165 | + }) |
| 166 | +} |
0 commit comments