Mirrors
/
hf_text-generation-inference
mirror of https://github.com/huggingface/text-generation-inference.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Adding more cards.

This commit is contained in:
Nicolas Patry 2024-11-05 09:47:55 +01:00
parent fa912440b1
commit 23c0a20dc9
No known key found for this signature in database
GPG Key ID: D2920555C90F704C
2 changed files with 12 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
launcher/src/main.rs
View File

@ -33,7 +33,7 @@ fn compute_optimal(config: Option<&Config>, compute: Option<&ComputeType>) -> Op
if let (Some(config), Some(compute)) = (config, compute) {
if let (Some(f16_max_compute), Some(model_compute)) = (compute.f16_flop(), config.flop()) {
tracing::debug!("MAx compute {f16_max_compute} model compute {model_compute}");
let optimal_size = (f16_max_compute / model_compute) as usize;
let optimal_size = (f16_max_compute / model_compute / 2) as usize;
if optimal_size > 100 {
// Ignore calculations that's too low
// Most likely an error
@ -1500,6 +1500,8 @@ impl ComputeType {
let card_flop = match &self.card[..] {
// https://www.nvidia.com/en-us/data-center/l4/
"nvidia-l4" => Some(121 * 10u64.pow(12)),
// https://www.nvidia.com/en-us/data-center/products/a10-gpu/
"nvidia-a10g" => Some(125 * 10u64.pow(12)),
card => {
tracing::warn!("Unkown compute for card {card}");
None

10
server/text_generation_server/models/flash_causal_lm.py
View File

@ -1554,12 +1554,16 @@ class FlashCausalLM(Model):
)
batch_num_blocks = batch.num_blocks
num_tokens = batch.to_pb().current_tokens
logger.info(f"BLOCKS {batch.num_blocks}")
free_memory = get_free_memory(self.device, MEMORY_FRACTION)
logger.info(f"Free memory {free_memory}")
if SYSTEM == "rocm" and os.environ.get("PYTORCH_TUNABLEOP_ENABLED", False):
torch.cuda.tunable.tuning_enable(False)
_, _batch, _ = self.generate_token(batch)
except torch.cuda.OutOfMemoryError as e:
raise RuntimeError(
f"Not enough memory to handle {batch.to_pb().current_tokens} prefill tokens. "
f"Not enough memory to handle {num_tokens} prefill tokens. "
f"You need to decrease `--max-batch-prefill-tokens`"
) from e
@ -2106,6 +2110,10 @@ class FlashCausalLM(Model):
if prefill and prefill_logprobs:
# Get prefill logprobs with inplace softmax (avoid copying the `out` tensor (max_batch_prefill_tokens * vocab_size))
free_memory = get_free_memory(self.device, MEMORY_FRACTION)
logger.info(f"Free memory {free_memory / 1e9}GB")
logmemory = out.nelement() * out.element_size()
logger.info(f"Log memory {logmemory / 1e9}GB")
torch.log_softmax(out, -1, out=out)
prefill_logprobs_tensor = out
prefill_logprobs = torch.gather(