huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-04-24 08:22:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update habana profiler (#50) (#80)

Browse Source 
Co-authored-by: mswiniarsk <156412439+mswiniarsk@users.noreply.github.com>
This commit is contained in:
Karol Damaszke 2024-02-28 09:57:40 +01:00 committed by GitHub
parent d31fb62576
commit 31bed905d4
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 24 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
README.md
View File

@ -74,8 +74,9 @@ Environment Variables Added:
| --------------------------- | :--------- | :--------------- | :------------------------------------------------------------------------------------------------------------------------------- | :--------------------------- | | --------------------------- | :--------- | :--------------- | :------------------------------------------------------------------------------------------------------------------------------- | :--------------------------- |
| MAX_TOTAL_TOKENS | integer | 0 | Control the padding of input | add -e in docker run, such | | MAX_TOTAL_TOKENS | integer | 0 | Control the padding of input | add -e in docker run, such |
| ENABLE_HPU_GRAPH | true/false | true | Enable hpu graph or not | add -e in docker run command | | ENABLE_HPU_GRAPH | true/false | true | Enable hpu graph or not | add -e in docker run command |
| PROF_WARMUPSTEP | integer | 0 | Enable/disable profile, control profile warmup step, 0 means disable profile | add -e in docker run command | | PROF_WAITSTEP | integer | 0 | Control profile wait steps | add -e in docker run command |
| PROF_STEP | integer | 5 | Control profile step | add -e in docker run command | | PROF_WARMUPSTEP | integer | 0 | Control profile warmup steps | add -e in docker run command |
| PROF_STEP | integer | 0 | Enable/disable profile, control profile active steps | add -e in docker run command |
| PROF_PATH | string | /tmp/hpu_profile | Define profile folder | add -e in docker run command | | PROF_PATH | string | /tmp/hpu_profile | Define profile folder | add -e in docker run command |
| PROF_RANKS | string | 0 | Comma-separated list of ranks to profile | add -e in docker run command | | PROF_RANKS | string | 0 | Comma-separated list of ranks to profile | add -e in docker run command |
| PROF_RECORD_SHAPES | true/false | false | Control record_shapes option in the profiler | add -e in docker run command | | PROF_RECORD_SHAPES | true/false | false | Control record_shapes option in the profiler | add -e in docker run command |

38
server/text_generation_server/models/causal_lm.py
View File

@ -324,7 +324,8 @@ class CausalLMBatch(Batch):
htorch.core.mark_step() htorch.core.mark_step()
dst_keys = [prepare_memory(new_bs * chunk_size, prev, inplace) for prev in src_keys[target_batch_idx]] dst_keys = [prepare_memory(new_bs * chunk_size, prev, inplace) for prev in src_keys[target_batch_idx]]
dst_keys = [move_data(dst_keys[layer_num], chunk_size, indices, [src[layer_num] for src in src_keys]) for layer_num in range(num_layers)] dst_keys = [move_data(dst_keys[layer_num], chunk_size, indices, [src[layer_num]
for src in src_keys]) for layer_num in range(num_layers)]
src_values = [torch.stack(src) for src in src_values] src_values = [torch.stack(src) for src in src_values]
htorch.core.mark_step() htorch.core.mark_step()
@ -334,7 +335,8 @@ class CausalLMBatch(Batch):
htorch.core.mark_step() htorch.core.mark_step()
dst_values = [prepare_memory(new_bs * chunk_size, prev, inplace) for prev in src_values[target_batch_idx]] dst_values = [prepare_memory(new_bs * chunk_size, prev, inplace) for prev in src_values[target_batch_idx]]
dst_values = [move_data(dst_values[layer_num], chunk_size, indices, [src[layer_num] for src in src_values]) for layer_num in range(num_layers)] dst_values = [move_data(dst_values[layer_num], chunk_size, indices, [src[layer_num]
for src in src_values]) for layer_num in range(num_layers)]
past_key_values = past_key_values_type(zip(dst_keys, dst_values)) past_key_values = past_key_values_type(zip(dst_keys, dst_values))
@ -626,18 +628,20 @@ class CausalLM(Model):
) )
prof_ranks = [int(val) for val in os.getenv("PROF_RANKS", "0").split(',')] prof_ranks = [int(val) for val in os.getenv("PROF_RANKS", "0").split(',')]
self.profiling_warmup_steps = int(os.getenv("PROF_WARMUPSTEP", "0")) if rank in prof_ranks else 0 self.profiling_warmup_steps = int(os.getenv("PROF_WARMUPSTEP", "0")) if rank in prof_ranks else 0
self.profiling_steps = int(os.getenv("PROF_STEP", "5")) self.profiling_steps = int(os.getenv("PROF_STEP", "0")) if rank in prof_ranks else 0
self.profiling_wait_steps = int(os.getenv("PROF_WAITSTEP", "0"))
record_shapes = os.getenv("PROF_RECORD_SHAPES", "false").lower() == "true" record_shapes = os.getenv("PROF_RECORD_SHAPES", "false").lower() == "true"
output_dir = os.getenv("PROF_PATH", "/tmp/hpu_profile") output_dir = os.getenv("PROF_PATH", "/tmp/hpu_profile")
self.hb_profer = HabanaProfile( if self.profiling_steps > 0:
warmup=self.profiling_warmup_steps, active=self.profiling_steps, output_dir=output_dir, record_shapes=record_shapes self.hb_profiler = HabanaProfile(
) wait=self.profiling_wait_steps,
if self.profiling_warmup_steps > 0: warmup=self.profiling_warmup_steps,
self.hb_profer_started = True active=self.profiling_steps,
self.hb_profer.start() output_dir=output_dir, record_shapes=record_shapes
)
self.hb_profiler.start()
else: else:
self.hb_profer = None self.hb_profiler = None
self.hb_profer_started = False
self.step = 0 self.step = 0
def setup_quantization(self, model): def setup_quantization(self, model):
@ -979,10 +983,10 @@ class CausalLM(Model):
req.prefix_offset = prefix_offset req.prefix_offset = prefix_offset
req.read_offset = read_offset req.read_offset = read_offset
htorch.core.mark_step() htorch.core.mark_step()
self.step = self.step + 1 self.step = self.step + 1
if self.hb_profer_started == True and self.step > self.profiling_warmup_steps + self.profiling_steps: if self.hb_profiler is not None:
self.hb_profer.stop() if self.step > self.profiling_wait_steps + self.profiling_warmup_steps + self.profiling_steps:
self.hb_profer_started = False self.hb_profiler.stop()
else:
return generations, batch if not stopped else None self.hb_profiler.step()
return generations, batch if not stopped else None