Moment infini #76
Open
Moment infini #76
Conversation
This file contains hidden or 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
added 2 commits
March 26, 2025 18:00
mononitogoswami
requested review from
JanekDev,
mononitogoswami and
zukowskanina
JanekDev
approved these changes
Mar 27, 2025
| """ | ||
|
|
||
| x_enc = self.tokenizer(x=x_enc) | ||
| batch_size, n_channels, seq_len = x_enc.shape |
There was a problem hiding this comment.
I suggest unifying n_channels and n_series.
momentfm/models/moment.py
Outdated
| x: [batch_size x n_channels x n_patches x d_model] | ||
| output: [batch_size x n_channels x forecast_horizon] | ||
| """ | ||
| x = self.flatten(x) # x: [batch_size, n_series, n_patches, d_model] |
There was a problem hiding this comment.
suggesting unification of n_channels and n_series
momentfm/utils/t5_infini.py
Outdated
| if not self.has_relative_attention_bias: | ||
| position_bias = torch.zeros( | ||
| (1, self.n_channels, self.n_heads, seq_length, key_length), device=hidden_states.device, dtype=hidden_states.dtype | ||
| ) # Willa - should we use n_channels or just 1? |
There was a problem hiding this comment.
in the original implementation by Nina there is no channel axis, so it gets probably broadcasted and position biases are shared between channels, hence there should be probably 1?
momentfm/utils/t5_infini.py
Outdated
| # Vectorized infini attention computation across channels | ||
| sigma_k = self.elu(key_states) + 1.0 # [batch_size, n_series, n_heads, n_patch, dim] | ||
| sigma_k_transposed = sigma_k.transpose(-2, -1) # [batch_size, n_series, n_heads, dim, n_patch] | ||
| memory_matrix = torch.matmul(sigma_k_transposed, value_states).sum(dim=1).unsqueeze(1) # [batch_size, 1, n_heads, dim, dim] sum over channels then unsqueeze to enable broadcasting over channels |
There was a problem hiding this comment.
for the purpose of making it easier to understand- can we split the computation of memory matrix into memory updates and only then sum them in the separate line? Implementation looks correct btw!
momentfm/utils/t5_infini.py
Outdated
| z = sigma_k.sum(dim=-2).unsqueeze(-1).sum(dim=1) # [batch_size, n_heads, dim, 1] sum over sequence length and channels | ||
| z = z.unsqueeze(dim=1) # [batch_size, 1, n_heads, dim, 1] | ||
| sigma_q = self.elu(query_states) + 1.0 # [batch_size, n_series, n_heads, n_patch, dim] | ||
| A_mem = (sigma_q @ memory_matrix) / ((sigma_q @ z) + 1e-6) # [batch_size, n_series, n_heads, n_patch, dim]/[batch_size, n_series, n_heads, n_patch, 1] --> [batch_size, n_series, n_heads, n_patch, dim] Adding 1e-6 for preventing division to 0 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
New files: