llama : (mrope) allow using normal 1D position for text token #13138
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
ngxson
commented
Apr 27, 2025
src/llama-graph.cpp
Outdated
| @@ -1031,11 +1042,12 @@ ggml_tensor * llm_graph_context::build_inp_pos() const { | |||
| } | |||
|
|
|||
| ggml_tensor * llm_graph_context::build_inp_attn_scale() const { | |||
| auto inp = std::make_unique<llm_graph_input_attn_temp>(n_pos_per_token(), hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale); | |||
| auto inp = std::make_unique<llm_graph_input_attn_temp>(n_pos_per_embd(), hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale); | |||
There was a problem hiding this comment.
@ggerganov Because build_inp_attn_scale is currently used exclusively by llama 4, do you think we should get rid of n_pos_per_embd and replace it with a GGML_ASSERT(n_pos_per_embd() == 1) ?
The main motivation is to make this code looks less complicated, as there is ~0% chance Qwen model gonna use this
There was a problem hiding this comment.
On second thought, build_inp_attn_scale should work well even in the case of N pos per token.
That's because the scale is applied per embedding, and the number of embedding is independent from N pos per token.
In any cases, I removed the n_pos_per_embd in 9cd16a3 , merging this PR once the CI is green
ggerganov
approved these changes
Apr 28, 2025
src/llama-graph.cpp
Outdated
| @@ -1031,11 +1042,12 @@ ggml_tensor * llm_graph_context::build_inp_pos() const { | |||
| } | |||
|
|
|||
| ggml_tensor * llm_graph_context::build_inp_attn_scale() const { | |||
| auto inp = std::make_unique<llm_graph_input_attn_temp>(n_pos_per_token(), hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale); | |||
| auto inp = std::make_unique<llm_graph_input_attn_temp>(n_pos_per_embd(), hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale); | |||
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.
For M-RoPE, we want to use normal 1D position for text token.
This is done to simplify the use case of
llama_decode()with text tokens, which is needed for adding Qwen2VL tolibmtmdand toserver.cppThis should also align with #11875, because in the future we want text position to be tracked internally by
libllama