Skip to content

spec: keep DFlash vision positions contiguous #4

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.

Sign up for GitHub

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

Open
Sphinx-8 wants to merge 1 commit into
base: dflash2
Choose a base branch
from
Open

spec: keep DFlash vision positions contiguous #4

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 17 additions & 2 deletions common/speculative.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1096,6 +1096,17 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {

const int32_t n_ubatch = (int32_t) llama_n_ubatch(ctx_dft);

// The target may use multi-dimensional position IDs for multimodal
// embeddings (for example Qwen-VL M-RoPE). The DFlash draft is a
// text-only decoder with a one-dimensional KV cache, so copying those
// repeated spatial positions makes its sequence non-consecutive.
// Keep target positions untouched and inject target features into the
// draft cache at strictly increasing per-sequence positions instead.
std::vector<llama_pos> next_inject_pos(n_seq, -1);
for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
next_inject_pos[seq_id] = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), seq_id) + 1;
}

// Flatten token-wise encoder work into shared chunks while preserving each row's position and sequence.
for (int32_t offset = 0; offset < n_tokens; offset += n_ubatch) {
const int32_t n_chunk = std::min(n_ubatch, n_tokens - offset);
Expand Down Expand Up @@ -1139,7 +1150,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
GGML_ASSERT(batch_in.n_seq_id[j] == 1);
const llama_seq_id seq_id = batch_in.seq_id[j][0];
GGML_ASSERT(seq_id >= 0 && seq_id < (llama_seq_id) n_seq);
batch_inject.pos[i] = batch_in.pos[j];
batch_inject.pos[i] = next_inject_pos[seq_id]++;
batch_inject.n_seq_id[i] = 1;
batch_inject.seq_id[i][0] = seq_id;
batch_inject.logits[i] = false;
Expand Down Expand Up @@ -1176,7 +1187,11 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {

common_sampler_reset(smpls[seq_id].get());

const int32_t n = (int32_t) dp.n_past;
// Multimodal target prompts can have many embedding rows while the
// text-only draft cache advances on its own 1-D position scale.
// Draft from the cache's actual next position instead of assuming
// target token count and draft position are identical.
const int32_t n = (int32_t) llama_memory_seq_pos_max(llama_get_memory(ctx_dft), seq_id) + 1;

const int32_t n_draft = params.n_max;

Expand Down