mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2024-12-29 12:54:35 +00:00
55e47786e3
* llama : deprecate softmax sampler + fix dist sampler ggml-ci * tests : replace macros with functions ggml-ci * sampling : change temperature sampler logic For t <= 0.0f, keep the max logit intact and set the rest to -inf * cont : no need for special "greedy" logic top-k == 1 is the same * tests : init prob correctly * llama : handle temp <= 0.0 in the temp_ext sampler too ggml-ci * cont : avoid extra loop in temperature sampler for sub-zero temp ggml-ci
336 lines
11 KiB
Swift
336 lines
11 KiB
Swift
import Foundation
|
|
import llama
|
|
|
|
enum LlamaError: Error {
|
|
case couldNotInitializeContext
|
|
}
|
|
|
|
func llama_batch_clear(_ batch: inout llama_batch) {
|
|
batch.n_tokens = 0
|
|
}
|
|
|
|
func llama_batch_add(_ batch: inout llama_batch, _ id: llama_token, _ pos: llama_pos, _ seq_ids: [llama_seq_id], _ logits: Bool) {
|
|
batch.token [Int(batch.n_tokens)] = id
|
|
batch.pos [Int(batch.n_tokens)] = pos
|
|
batch.n_seq_id[Int(batch.n_tokens)] = Int32(seq_ids.count)
|
|
for i in 0..<seq_ids.count {
|
|
batch.seq_id[Int(batch.n_tokens)]![Int(i)] = seq_ids[i]
|
|
}
|
|
batch.logits [Int(batch.n_tokens)] = logits ? 1 : 0
|
|
|
|
batch.n_tokens += 1
|
|
}
|
|
|
|
actor LlamaContext {
|
|
private var model: OpaquePointer
|
|
private var context: OpaquePointer
|
|
private var sampling: UnsafeMutablePointer<llama_sampler>
|
|
private var batch: llama_batch
|
|
private var tokens_list: [llama_token]
|
|
var is_done: Bool = false
|
|
|
|
/// This variable is used to store temporarily invalid cchars
|
|
private var temporary_invalid_cchars: [CChar]
|
|
|
|
var n_len: Int32 = 1024
|
|
var n_cur: Int32 = 0
|
|
|
|
var n_decode: Int32 = 0
|
|
|
|
init(model: OpaquePointer, context: OpaquePointer) {
|
|
self.model = model
|
|
self.context = context
|
|
self.tokens_list = []
|
|
self.batch = llama_batch_init(512, 0, 1)
|
|
self.temporary_invalid_cchars = []
|
|
let sparams = llama_sampler_chain_default_params()
|
|
self.sampling = llama_sampler_chain_init(sparams)
|
|
llama_sampler_chain_add(self.sampling, llama_sampler_init_temp(0.4))
|
|
llama_sampler_chain_add(self.sampling, llama_sampler_init_dist(1234))
|
|
}
|
|
|
|
deinit {
|
|
llama_sampler_free(sampling)
|
|
llama_batch_free(batch)
|
|
llama_free(context)
|
|
llama_free_model(model)
|
|
llama_backend_free()
|
|
}
|
|
|
|
static func create_context(path: String) throws -> LlamaContext {
|
|
llama_backend_init()
|
|
var model_params = llama_model_default_params()
|
|
|
|
#if targetEnvironment(simulator)
|
|
model_params.n_gpu_layers = 0
|
|
print("Running on simulator, force use n_gpu_layers = 0")
|
|
#endif
|
|
let model = llama_load_model_from_file(path, model_params)
|
|
guard let model else {
|
|
print("Could not load model at \(path)")
|
|
throw LlamaError.couldNotInitializeContext
|
|
}
|
|
|
|
let n_threads = max(1, min(8, ProcessInfo.processInfo.processorCount - 2))
|
|
print("Using \(n_threads) threads")
|
|
|
|
var ctx_params = llama_context_default_params()
|
|
ctx_params.n_ctx = 2048
|
|
ctx_params.n_threads = Int32(n_threads)
|
|
ctx_params.n_threads_batch = Int32(n_threads)
|
|
|
|
let context = llama_new_context_with_model(model, ctx_params)
|
|
guard let context else {
|
|
print("Could not load context!")
|
|
throw LlamaError.couldNotInitializeContext
|
|
}
|
|
|
|
return LlamaContext(model: model, context: context)
|
|
}
|
|
|
|
func model_info() -> String {
|
|
let result = UnsafeMutablePointer<Int8>.allocate(capacity: 256)
|
|
result.initialize(repeating: Int8(0), count: 256)
|
|
defer {
|
|
result.deallocate()
|
|
}
|
|
|
|
// TODO: this is probably very stupid way to get the string from C
|
|
|
|
let nChars = llama_model_desc(model, result, 256)
|
|
let bufferPointer = UnsafeBufferPointer(start: result, count: Int(nChars))
|
|
|
|
var SwiftString = ""
|
|
for char in bufferPointer {
|
|
SwiftString.append(Character(UnicodeScalar(UInt8(char))))
|
|
}
|
|
|
|
return SwiftString
|
|
}
|
|
|
|
func get_n_tokens() -> Int32 {
|
|
return batch.n_tokens;
|
|
}
|
|
|
|
func completion_init(text: String) {
|
|
print("attempting to complete \"\(text)\"")
|
|
|
|
tokens_list = tokenize(text: text, add_bos: true)
|
|
temporary_invalid_cchars = []
|
|
|
|
let n_ctx = llama_n_ctx(context)
|
|
let n_kv_req = tokens_list.count + (Int(n_len) - tokens_list.count)
|
|
|
|
print("\n n_len = \(n_len), n_ctx = \(n_ctx), n_kv_req = \(n_kv_req)")
|
|
|
|
if n_kv_req > n_ctx {
|
|
print("error: n_kv_req > n_ctx, the required KV cache size is not big enough")
|
|
}
|
|
|
|
for id in tokens_list {
|
|
print(String(cString: token_to_piece(token: id) + [0]))
|
|
}
|
|
|
|
llama_batch_clear(&batch)
|
|
|
|
for i1 in 0..<tokens_list.count {
|
|
let i = Int(i1)
|
|
llama_batch_add(&batch, tokens_list[i], Int32(i), [0], false)
|
|
}
|
|
batch.logits[Int(batch.n_tokens) - 1] = 1 // true
|
|
|
|
if llama_decode(context, batch) != 0 {
|
|
print("llama_decode() failed")
|
|
}
|
|
|
|
n_cur = batch.n_tokens
|
|
}
|
|
|
|
func completion_loop() -> String {
|
|
var new_token_id: llama_token = 0
|
|
|
|
new_token_id = llama_sampler_sample(sampling, context, batch.n_tokens - 1)
|
|
|
|
if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
|
|
print("\n")
|
|
is_done = true
|
|
let new_token_str = String(cString: temporary_invalid_cchars + [0])
|
|
temporary_invalid_cchars.removeAll()
|
|
return new_token_str
|
|
}
|
|
|
|
let new_token_cchars = token_to_piece(token: new_token_id)
|
|
temporary_invalid_cchars.append(contentsOf: new_token_cchars)
|
|
let new_token_str: String
|
|
if let string = String(validatingUTF8: temporary_invalid_cchars + [0]) {
|
|
temporary_invalid_cchars.removeAll()
|
|
new_token_str = string
|
|
} else if (0 ..< temporary_invalid_cchars.count).contains(where: {$0 != 0 && String(validatingUTF8: Array(temporary_invalid_cchars.suffix($0)) + [0]) != nil}) {
|
|
// in this case, at least the suffix of the temporary_invalid_cchars can be interpreted as UTF8 string
|
|
let string = String(cString: temporary_invalid_cchars + [0])
|
|
temporary_invalid_cchars.removeAll()
|
|
new_token_str = string
|
|
} else {
|
|
new_token_str = ""
|
|
}
|
|
print(new_token_str)
|
|
// tokens_list.append(new_token_id)
|
|
|
|
llama_batch_clear(&batch)
|
|
llama_batch_add(&batch, new_token_id, n_cur, [0], true)
|
|
|
|
n_decode += 1
|
|
n_cur += 1
|
|
|
|
if llama_decode(context, batch) != 0 {
|
|
print("failed to evaluate llama!")
|
|
}
|
|
|
|
return new_token_str
|
|
}
|
|
|
|
func bench(pp: Int, tg: Int, pl: Int, nr: Int = 1) -> String {
|
|
var pp_avg: Double = 0
|
|
var tg_avg: Double = 0
|
|
|
|
var pp_std: Double = 0
|
|
var tg_std: Double = 0
|
|
|
|
for _ in 0..<nr {
|
|
// bench prompt processing
|
|
|
|
llama_batch_clear(&batch)
|
|
|
|
let n_tokens = pp
|
|
|
|
for i in 0..<n_tokens {
|
|
llama_batch_add(&batch, 0, Int32(i), [0], false)
|
|
}
|
|
batch.logits[Int(batch.n_tokens) - 1] = 1 // true
|
|
|
|
llama_kv_cache_clear(context)
|
|
|
|
let t_pp_start = ggml_time_us()
|
|
|
|
if llama_decode(context, batch) != 0 {
|
|
print("llama_decode() failed during prompt")
|
|
}
|
|
llama_synchronize(context)
|
|
|
|
let t_pp_end = ggml_time_us()
|
|
|
|
// bench text generation
|
|
|
|
llama_kv_cache_clear(context)
|
|
|
|
let t_tg_start = ggml_time_us()
|
|
|
|
for i in 0..<tg {
|
|
llama_batch_clear(&batch)
|
|
|
|
for j in 0..<pl {
|
|
llama_batch_add(&batch, 0, Int32(i), [Int32(j)], true)
|
|
}
|
|
|
|
if llama_decode(context, batch) != 0 {
|
|
print("llama_decode() failed during text generation")
|
|
}
|
|
llama_synchronize(context)
|
|
}
|
|
|
|
let t_tg_end = ggml_time_us()
|
|
|
|
llama_kv_cache_clear(context)
|
|
|
|
let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
|
|
let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
|
|
|
|
let speed_pp = Double(pp) / t_pp
|
|
let speed_tg = Double(pl*tg) / t_tg
|
|
|
|
pp_avg += speed_pp
|
|
tg_avg += speed_tg
|
|
|
|
pp_std += speed_pp * speed_pp
|
|
tg_std += speed_tg * speed_tg
|
|
|
|
print("pp \(speed_pp) t/s, tg \(speed_tg) t/s")
|
|
}
|
|
|
|
pp_avg /= Double(nr)
|
|
tg_avg /= Double(nr)
|
|
|
|
if nr > 1 {
|
|
pp_std = sqrt(pp_std / Double(nr - 1) - pp_avg * pp_avg * Double(nr) / Double(nr - 1))
|
|
tg_std = sqrt(tg_std / Double(nr - 1) - tg_avg * tg_avg * Double(nr) / Double(nr - 1))
|
|
} else {
|
|
pp_std = 0
|
|
tg_std = 0
|
|
}
|
|
|
|
let model_desc = model_info();
|
|
let model_size = String(format: "%.2f GiB", Double(llama_model_size(model)) / 1024.0 / 1024.0 / 1024.0);
|
|
let model_n_params = String(format: "%.2f B", Double(llama_model_n_params(model)) / 1e9);
|
|
let backend = "Metal";
|
|
let pp_avg_str = String(format: "%.2f", pp_avg);
|
|
let tg_avg_str = String(format: "%.2f", tg_avg);
|
|
let pp_std_str = String(format: "%.2f", pp_std);
|
|
let tg_std_str = String(format: "%.2f", tg_std);
|
|
|
|
var result = ""
|
|
|
|
result += String("| model | size | params | backend | test | t/s |\n")
|
|
result += String("| --- | --- | --- | --- | --- | --- |\n")
|
|
result += String("| \(model_desc) | \(model_size) | \(model_n_params) | \(backend) | pp \(pp) | \(pp_avg_str) ± \(pp_std_str) |\n")
|
|
result += String("| \(model_desc) | \(model_size) | \(model_n_params) | \(backend) | tg \(tg) | \(tg_avg_str) ± \(tg_std_str) |\n")
|
|
|
|
return result;
|
|
}
|
|
|
|
func clear() {
|
|
tokens_list.removeAll()
|
|
temporary_invalid_cchars.removeAll()
|
|
llama_kv_cache_clear(context)
|
|
}
|
|
|
|
private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
|
|
let utf8Count = text.utf8.count
|
|
let n_tokens = utf8Count + (add_bos ? 1 : 0) + 1
|
|
let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
|
|
let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, false)
|
|
|
|
var swiftTokens: [llama_token] = []
|
|
for i in 0..<tokenCount {
|
|
swiftTokens.append(tokens[Int(i)])
|
|
}
|
|
|
|
tokens.deallocate()
|
|
|
|
return swiftTokens
|
|
}
|
|
|
|
/// - note: The result does not contain null-terminator
|
|
private func token_to_piece(token: llama_token) -> [CChar] {
|
|
let result = UnsafeMutablePointer<Int8>.allocate(capacity: 8)
|
|
result.initialize(repeating: Int8(0), count: 8)
|
|
defer {
|
|
result.deallocate()
|
|
}
|
|
let nTokens = llama_token_to_piece(model, token, result, 8, 0, false)
|
|
|
|
if nTokens < 0 {
|
|
let newResult = UnsafeMutablePointer<Int8>.allocate(capacity: Int(-nTokens))
|
|
newResult.initialize(repeating: Int8(0), count: Int(-nTokens))
|
|
defer {
|
|
newResult.deallocate()
|
|
}
|
|
let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens, 0, false)
|
|
let bufferPointer = UnsafeBufferPointer(start: newResult, count: Int(nNewTokens))
|
|
return Array(bufferPointer)
|
|
} else {
|
|
let bufferPointer = UnsafeBufferPointer(start: result, count: Int(nTokens))
|
|
return Array(bufferPointer)
|
|
}
|
|
}
|
|
}
|