model : add LiquidAI/d1-2B and LiquidAI/d1-omni-2B (#31275)
* model : add LiquidAI/d1-2B and LiquidAI/d1-omni-2B Assisted-by: Claude Opus 5.5 * convert : keep the speech task templates out of the d1 GGUFs Assisted-by: Claude Opus 5.5 * small refactoring: no_batching --------- Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
This commit is contained in:
1 parent
1b4554eff1
commit
d05706b4f3
19 files changed
+307
-63
No files matched your search
@@ -529,6 +529,12 @@ struct server_slot {
|
||||
return false;
|
||||
}
|
||||
|
||||
// lfm3 d1 has no memory, its prompt is read as one flat ubatch and the short convs would mix the sequences
|
||||
// TODO: batch them along the sequence dimension, this needs ubatches of sequences with unequal lengths
|
||||
if (task->decision.no_batching || other_slot.task->decision.no_batching) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return task->type == other_slot.task->type
|
||||
&& inp_embd.size() == other_slot.inp_embd.size()
|
||||
&& are_lora_equal(lora, other_slot.lora);
|
||||
@@ -3955,7 +3961,8 @@ private:
|
||||
}
|
||||
|
||||
// TODO: support memory-less logits computation
|
||||
if (slot.task->need_logits() && !llama_get_memory(ctx_tgt)) {
|
||||
// a decision reads its logits from the one batch of its prompt
|
||||
if (slot.task->need_logits() && !llama_get_memory(ctx_tgt) && slot.task->type != SERVER_TASK_TYPE_DECISION) {
|
||||
send_error(slot, "the current context does not support logits computation. skipping", ERROR_TYPE_SERVER);
|
||||
slot.release();
|
||||
return;
|
||||
|
||||
Reference in new issue
Block a user