Skip to content

llama : grouped-query attention + LLaMAv2 70B support #2276

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Jul 23, 2023
Merged

llama : grouped-query attention + LLaMAv2 70B support #2276

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
c1893cd
CUDA: GQA implementation
JohannesGaessler Jul 22, 2023
3fdc00f
llama : support for GQA and LLaMAv2 70B
ggerganov Jul 23, 2023
2dac31b
py : fix hparams parsing (if-else blocks)
ggerganov Jul 23, 2023
c594992
py : oh boy ..
ggerganov Jul 23, 2023
8194d59
help : fix gqa value for 70B
ggerganov Jul 23, 2023
3ed2553
Merge branch 'master' into llama-v2-70b
ggerganov Jul 23, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
66 changes: 45 additions & 21 deletions convert.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -142,9 +142,9 @@ def find_n_mult(n_ff: int, n_embd: int) -> int:
@dataclass
class Params:
n_vocab: int
n_embd: int
n_mult: int
n_head: int
n_embd: int
n_mult: int
n_head: int
n_layer: int

@staticmethod
Expand All @@ -167,40 +167,65 @@ def guessed(model: 'LazyModel') -> 'Params':
n_head=n_embd // 128 # guessed

return Params(
n_vocab=n_vocab,
n_embd=n_embd,
n_mult=256,
n_head=n_head,
n_layer=n_layer,
n_vocab = n_vocab,
n_embd = n_embd,
n_mult = 256,
n_head = n_head,
n_layer = n_layer,
)

@staticmethod
def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
config = json.load(open(config_path))

n_vocab = config["vocab_size"];
n_embd = config["hidden_size"];
n_head = config["num_attention_heads"];
n_embd = config["hidden_size"];
n_head = config["num_attention_heads"];
n_layer = config["num_hidden_layers"];
n_ff = config["intermediate_size"];
n_ff = config["intermediate_size"];

n_mult = find_n_mult(n_ff, n_embd);

return Params(
n_vocab=n_vocab,
n_embd=n_embd,
n_mult=n_mult,
n_head=n_head,
n_layer=n_layer,
n_vocab = n_vocab,
n_embd = n_embd,
n_mult = n_mult,
n_head = n_head,
n_layer = n_layer,
)

# LLaMA v2 70B params.json
# {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1
@staticmethod
def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
config = json.load(open(config_path))

n_vocab = config["vocab_size"];
n_embd = config["dim"];
n_head = config["n_heads"];
n_layer = config["n_layers"];
n_mult = config["multiple_of"];

if n_vocab == -1:
n_vocab = model["tok_embeddings.weight"].shape[0]

return Params(
n_vocab = n_vocab,
n_embd = n_embd,
n_mult = n_mult,
n_head = n_head,
n_layer = n_layer,
)

@staticmethod
def load(model_plus: 'ModelPlus') -> 'Params':
hf_config_path = model_plus.paths[0].parent / "config.json"
orig_config_path = model_plus.paths[0].parent / "params.json"
hf_transformer_config_path = model_plus.paths[0].parent / "config.json"

if hf_transformer_config_path.exists():
params = Params.loadHFTransformerJson(model_plus.model, hf_transformer_config_path)
if hf_config_path.exists():
params = Params.loadHFTransformerJson(model_plus.model, hf_config_path)
elif orig_config_path.exists():
params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path)
else:
params = Params.guessed(model_plus.model)

Expand Down Expand Up @@ -1036,8 +1061,7 @@ def write_vocab(self, vocab: Vocab) -> None:
@staticmethod
def write_vocab_only(fname_out: Path, vocab: Vocab) -> None:
of = OutputFile(fname_out)
params = Params(n_vocab=vocab.vocab_size, n_embd=0, n_mult=0,
n_head=1, n_layer=0)
params = Params(n_vocab=vocab.vocab_size, n_embd=0, n_mult=0, n_head=1, n_layer=0)
of = OutputFile(fname_out)
of.write_file_header(params, file_type=GGMLFileType.AllF32)
of.write_vocab(vocab)
Expand Down
12 changes: 10 additions & 2 deletions examples/common.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -168,6 +168,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
break;
}
params.n_ctx = std::stoi(argv[i]);
} else if (arg == "-gqa" || arg == "--gqa") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.n_gqa = std::stoi(argv[i]);
} else if (arg == "--rope-freq-base") {
if (++i >= argc) {
invalid_param = true;
Expand Down Expand Up @@ -485,6 +491,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stdout, " -f FNAME, --file FNAME\n");
fprintf(stdout, " prompt file to start generation.\n");
fprintf(stdout, " -n N, --n-predict N number of tokens to predict (default: %d, -1 = infinity)\n", params.n_predict);
fprintf(stdout, " -c N, --ctx-size N size of the prompt context (default: %d)\n", params.n_ctx);
fprintf(stdout, " -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch);
fprintf(stdout, " -gqa N, --gqa N grouped-query attention factor (TEMP!!! use 8 for LLaMAv2 70B) (default: %d)\n", params.n_gqa);
fprintf(stdout, " --top-k N top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
fprintf(stdout, " --top-p N top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
fprintf(stdout, " --tfs N tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
Expand All @@ -505,15 +514,13 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stdout, " --cfg-negative-prompt PROMPT \n");
fprintf(stdout, " negative prompt to use for guidance. (default: empty)\n");
fprintf(stdout, " --cfg-scale N strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
fprintf(stdout, " -c N, --ctx-size N size of the prompt context (default: %d)\n", params.n_ctx);
fprintf(stdout, " --rope-freq-base N RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
fprintf(stdout, " --rope-freq-scale N RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
fprintf(stdout, " --ignore-eos ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
fprintf(stdout, " --no-penalize-nl do not penalize newline token\n");
fprintf(stdout, " --memory-f32 use f32 instead of f16 for memory key+value (default: disabled)\n");
fprintf(stdout, " not recommended: doubles context memory required and no measurable increase in quality\n");
fprintf(stdout, " --temp N temperature (default: %.1f)\n", (double)params.temp);
fprintf(stdout, " -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch);
fprintf(stdout, " --perplexity compute perplexity over each ctx window of the prompt\n");
fprintf(stdout, " --perplexity-lines compute perplexity over each line of the prompt\n");
fprintf(stdout, " --keep number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
Expand Down Expand Up @@ -580,6 +587,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param

lparams.n_ctx = params.n_ctx;
lparams.n_batch = params.n_batch;
lparams.n_gqa = params.n_gqa;
lparams.n_gpu_layers = params.n_gpu_layers;
lparams.main_gpu = params.main_gpu;
lparams.tensor_split = params.tensor_split;
Expand Down
3 changes: 2 additions & 1 deletion examples/common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,7 @@ struct gpt_params {
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 512; // context size
int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_gqa = 1; // grouped-query attention factor (TODO: move to hparams)
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited)
int32_t n_gpu_layers = 0; // number of layers to store in VRAM
Expand All @@ -47,7 +48,7 @@ struct gpt_params {
int32_t repeat_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float frequency_penalty = 0.00f; // 0.0 = disabled
float presence_penalty = 0.00f; // 0.0 = disabled
int mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate

Expand Down
4 changes: 2 additions & 2 deletions examples/main/main.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,8 +93,8 @@ int main(int argc, char ** argv) {
}

if (params.n_ctx > 2048) {
fprintf(stderr, "%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified);"
" you are on your own\n", __func__, params.n_ctx);
// TODO: determine the actual max context of the model (e.g. 4096 for LLaMA v2) and use that instead of 2048
fprintf(stderr, "%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified)\n", __func__, params.n_ctx);
} else if (params.n_ctx < 8) {
fprintf(stderr, "%s: warning: minimum context size is 8, using minimum size.\n", __func__);
params.n_ctx = 8;
Expand Down
Loading