Update flash_attention_fwd_benchmark.py
#2265
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Anatoly Myachev <[email protected]>
anmyachev
force-pushed
the
anmyachev-patch-1
branch
from
80a8f26
to
92998b2
Compare
Signed-off-by: Anatoly Myachev <[email protected]>
Signed-off-by: Anatoly Myachev <[email protected]>
Signed-off-by: Anatoly Myachev <[email protected]>
This reverts commit de9335c.
Signed-off-by: Anatoly Myachev <[email protected]>
anmyachev
commented
Sep 19, 2024
| q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False, scale=sm_scale).to(torch.float32) | ||
| atol = 1e-1 if N_CTX == 16384 else 1e-2 | ||
| benchmark_suit.assert_close(triton_fn(), torch_fn(), atol=atol, rtol=1e-3, err_msg='triton to torch') | ||
| torch_fn = lambda: torch.nn.functional.scaled_dot_product_attention( |
There was a problem hiding this comment.
Using ZE_FLAT_DEVICE_HIERARCHY=COMPOSITE the available memory is doubled and there is no more out of memory error for upstream pytorch (however, this affects the performance)
Signed-off-by: Anatoly Myachev <[email protected]>
anmyachev
commented
Sep 20, 2024
| @@ -64,7 +64,10 @@ def do_bench_ipex(fn, warmup=25, rep=100, grad_to_none=None, quantiles=None, fas | |||
| # We maintain a buffer of 256 MB that we clear | |||
| # before each kernel call to make sure that the L2 | |||
| # doesn't contain any input data before the run | |||
| cache_size = 256 * 1024 * 1024 | |||
| factor = 1 | |||
| if os.getenv("ZE_FLAT_DEVICE_HIERARCHY", "FLAT") == "COMPOSITE": | |||
There was a problem hiding this comment.
By increasing the cache for cleaning accordingly, the performance becomes +- the same for both cases
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
CI:
Error:
torch.OutOfMemoryError: XPU out of memory. Tried to allocate 32.00 GiB. GPU 0 has a total capacity of 64.00 GiB. Of the allocated memory 32.81 GiB is allocated by PyTorch, and 0 bytes is reserved by PyTorch but unallocated. Please use `empty_cache` to release all unoccupied cached memory.It's strange that
a total capacityis 64.00 GiB. I need to understand why (the expected capacity should be more in my understanding).UPD: Maybe it's related to https://spec.oneapi.io/level-zero/latest/core/PROG.html#environment-variables ZE_FLAT_DEVICE_HIERARCHY
ZE_FLAT_DEVICE_HIERARCHY=COMPOSITEcan help with this, but for now it has been decided to leave it as is