-
Notifications
You must be signed in to change notification settings - Fork 27
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.
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
Examples of training bias #84
Open
drisspg
wants to merge
1
commit into
main
Choose a base branch
from
drisspg/stack/2
base: main
Could not load branches
Branch not found: {{ refName }}
Loading
Could not load tags
Nothing to show
Loading
Are you sure you want to change the base?
Some commits from the old base branch may be removed from the timeline,
and old review comments may become outdated.
Open
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
drisspg
added a commit
that referenced
this pull request
Nov 22, 2024
stack-info: PR: #84, branch: drisspg/stack/2
drisspg
force-pushed
the
drisspg/stack/2
branch
from
November 22, 2024 22:06
d2864f6
to
0166927
Compare
facebook-github-bot
added
the
CLA Signed
This label is managed by the Meta Open Source bot.
label
Nov 22, 2024
stack-info: PR: #84, branch: drisspg/stack/2
drisspg
force-pushed
the
drisspg/stack/2
branch
from
November 22, 2024 22:07
0166927
to
4d045e1
Compare
pytorchmergebot
pushed a commit
to pytorch/pytorch
that referenced
this pull request
Nov 25, 2024
# Summary The follow up PR to: #137526. In this pr, we actually update the lowerings for the flex_attention backwards kernel to generate fused backward gradient calculations for any captured buffers that require grads. We are doing this using tl.atomic_add to scatter the correct gradients into zeroed out buffer for any captured buffers that required grads. Added many test cases and found. Along the way found some masking bugs. There are likely some performance cliffs here, specifically with D-types and on different GPUs. Planned to do this in a follow-up and profile the current strategy. We are explicitly choosing reduced memory over increased performance right now. By using atomics, we do not need to realize a full attention scores matrix. However, this comes with two downsides. One, this is potentially slower in some cases, and two, the gradient calculation for any captured buffers is non-deterministic. ## Worked Example Lets do the case where you are reading from one bias that doesn't require grad and using this to index into another that does. ScoreMod: ```Python bias = torch.randn( params.seq_length, device=self.device, dtype=params.dtype, requires_grad=True, ) offset = torch.randint( 0, params.seq_length, (params.seq_length,), device=self.device, ) def score_mod(score, b, h, q_idx, kv_idx): return score + bias[offset[q_idx]] ``` I am removing all but the new subgraph injected into the backwards: ``` Python dsT = pT * (dpT - Di[None, :]) # ~~~~~~~~~~~~~~~~~~~ Apply joint modification ~~~~~~~~~~~~~~~~~~~ grad_scores = (dsT) # ~~~~~~~~~~~~~~~~~~~ Apply other buffer grad writes ~~~~~~~~~~~~~ idx_b = off_z idx_h = off_hq idx_m = m idx_n = n scatter_mask = offs_m1[None, :] < Q_LEN and offs_n1[:, None] < KV_LEN tmp4 = (dsT).to(tl.float32) tl.atomic_add(out_ptr1 + (tl.broadcast_to(tl.load(in_ptr16 + idx_m), tmp4.shape)), tmp4, scatter_mask, sem='relaxed') # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ``` ## Key points * We always accumulate to float 32 grad buffers regardless of the type in the forward. This is because we normally do all computation intra kernel w/ fp32 accumulation and we want the same behavior for atomic additions * We are currently restricted to 1 scatter in the kenrel. I have some ideas on fx rewrites that would remove this restrictions but for now have nice error message w/ work around and will leave as a follow up. * Will do more extensive performance/ memory profiling in a follow up. ### Toy E2E example I have a toy E2E training example PR in the gym for now: pytorch-labs/attention-gym#84 I plan to update to a realistic learnable bias before landing Pull Request resolved: #137452 Approved by: https://github.com/Chillee
Ryo-not-rio
pushed a commit
to Ryo-not-rio/pytorch
that referenced
this pull request
Dec 2, 2024
…#137452) # Summary The follow up PR to: pytorch#137526. In this pr, we actually update the lowerings for the flex_attention backwards kernel to generate fused backward gradient calculations for any captured buffers that require grads. We are doing this using tl.atomic_add to scatter the correct gradients into zeroed out buffer for any captured buffers that required grads. Added many test cases and found. Along the way found some masking bugs. There are likely some performance cliffs here, specifically with D-types and on different GPUs. Planned to do this in a follow-up and profile the current strategy. We are explicitly choosing reduced memory over increased performance right now. By using atomics, we do not need to realize a full attention scores matrix. However, this comes with two downsides. One, this is potentially slower in some cases, and two, the gradient calculation for any captured buffers is non-deterministic. ## Worked Example Lets do the case where you are reading from one bias that doesn't require grad and using this to index into another that does. ScoreMod: ```Python bias = torch.randn( params.seq_length, device=self.device, dtype=params.dtype, requires_grad=True, ) offset = torch.randint( 0, params.seq_length, (params.seq_length,), device=self.device, ) def score_mod(score, b, h, q_idx, kv_idx): return score + bias[offset[q_idx]] ``` I am removing all but the new subgraph injected into the backwards: ``` Python dsT = pT * (dpT - Di[None, :]) # ~~~~~~~~~~~~~~~~~~~ Apply joint modification ~~~~~~~~~~~~~~~~~~~ grad_scores = (dsT) # ~~~~~~~~~~~~~~~~~~~ Apply other buffer grad writes ~~~~~~~~~~~~~ idx_b = off_z idx_h = off_hq idx_m = m idx_n = n scatter_mask = offs_m1[None, :] < Q_LEN and offs_n1[:, None] < KV_LEN tmp4 = (dsT).to(tl.float32) tl.atomic_add(out_ptr1 + (tl.broadcast_to(tl.load(in_ptr16 + idx_m), tmp4.shape)), tmp4, scatter_mask, sem='relaxed') # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ``` ## Key points * We always accumulate to float 32 grad buffers regardless of the type in the forward. This is because we normally do all computation intra kernel w/ fp32 accumulation and we want the same behavior for atomic additions * We are currently restricted to 1 scatter in the kenrel. I have some ideas on fx rewrites that would remove this restrictions but for now have nice error message w/ work around and will leave as a follow up. * Will do more extensive performance/ memory profiling in a follow up. ### Toy E2E example I have a toy E2E training example PR in the gym for now: pytorch-labs/attention-gym#84 I plan to update to a realistic learnable bias before landing Pull Request resolved: pytorch#137452 Approved by: https://github.com/Chillee
pobin6
pushed a commit
to pobin6/pytorch
that referenced
this pull request
Dec 5, 2024
…#137452) # Summary The follow up PR to: pytorch#137526. In this pr, we actually update the lowerings for the flex_attention backwards kernel to generate fused backward gradient calculations for any captured buffers that require grads. We are doing this using tl.atomic_add to scatter the correct gradients into zeroed out buffer for any captured buffers that required grads. Added many test cases and found. Along the way found some masking bugs. There are likely some performance cliffs here, specifically with D-types and on different GPUs. Planned to do this in a follow-up and profile the current strategy. We are explicitly choosing reduced memory over increased performance right now. By using atomics, we do not need to realize a full attention scores matrix. However, this comes with two downsides. One, this is potentially slower in some cases, and two, the gradient calculation for any captured buffers is non-deterministic. ## Worked Example Lets do the case where you are reading from one bias that doesn't require grad and using this to index into another that does. ScoreMod: ```Python bias = torch.randn( params.seq_length, device=self.device, dtype=params.dtype, requires_grad=True, ) offset = torch.randint( 0, params.seq_length, (params.seq_length,), device=self.device, ) def score_mod(score, b, h, q_idx, kv_idx): return score + bias[offset[q_idx]] ``` I am removing all but the new subgraph injected into the backwards: ``` Python dsT = pT * (dpT - Di[None, :]) # ~~~~~~~~~~~~~~~~~~~ Apply joint modification ~~~~~~~~~~~~~~~~~~~ grad_scores = (dsT) # ~~~~~~~~~~~~~~~~~~~ Apply other buffer grad writes ~~~~~~~~~~~~~ idx_b = off_z idx_h = off_hq idx_m = m idx_n = n scatter_mask = offs_m1[None, :] < Q_LEN and offs_n1[:, None] < KV_LEN tmp4 = (dsT).to(tl.float32) tl.atomic_add(out_ptr1 + (tl.broadcast_to(tl.load(in_ptr16 + idx_m), tmp4.shape)), tmp4, scatter_mask, sem='relaxed') # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ``` ## Key points * We always accumulate to float 32 grad buffers regardless of the type in the forward. This is because we normally do all computation intra kernel w/ fp32 accumulation and we want the same behavior for atomic additions * We are currently restricted to 1 scatter in the kenrel. I have some ideas on fx rewrites that would remove this restrictions but for now have nice error message w/ work around and will leave as a follow up. * Will do more extensive performance/ memory profiling in a follow up. ### Toy E2E example I have a toy E2E training example PR in the gym for now: pytorch-labs/attention-gym#84 I plan to update to a realistic learnable bias before landing Pull Request resolved: pytorch#137452 Approved by: https://github.com/Chillee
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.
Stacked PRs:
Examples of training bias