fix: chunked BF16, buffer cap, drop fake FMA#55
Merged
Conversation
1. mul_add with zero addend → plain multiply (was wasting an FMA slot) 2. Chunked row-batch reading for BF16 tensors: caps buffer at 128 MB regardless of tensor size. A 10.7 GB ffn_gate_exps reads in ~4.8K row batches instead of one 10.7 GB allocation. Minimum batch = 8 rows (one F64x8 SIMD width). 3. Buffer shrink_to after oversized tensors: bf16_buf is truncated back to MAX_BUF_ELEMS (64M u16 = 128 MB) if it somehow grew past. 4. Progress logging within large tensors: prints row count every chunk so you see activity during multi-minute tensor reads. read_tensor_bf16_raw() is now unused in the main path (kept for potential direct use in tests or smaller models).
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Addresses review items from the BF16-direct rebase:
Fix 1:
mul_addwith zero addendsums[bin].mul_add(splat(scale), splat(0.0))→sums[bin] * splat(scale)FMA with zero addend wastes the fuse — same latency as plain multiply but occupies the FMA port instead of the multiply port.
Fix 2+3: Chunked row-batch reading + buffer cap
Before:
read_tensor_bf16_rawallocates full tensor asVec<u16>. Forffn_gate_exps(128 experts × 5120 × 13824) that's 10.7 GB.After: Read in row batches capped at
MAX_BUF_ELEMS(64M u16 = 128 MB). Each batch: read → project → extend results. Buffer never exceeds 128 MB regardless of tensor size.142 chunks for the largest tensor, each fully processed before the next read. Peak RAM stays at ~128 MB instead of 10.7 GB.
shrink_to(MAX_BUF_ELEMS)after oversized tensors prevents the buffer from persisting at inflated size.Bonus: progress logging
Large tensors now print
... 4629/655360 rows (0.7%)per chunk so you see activity during multi-minute reads.What's unused
read_tensor_bf16_raw()is no longer called instream_index_gguf_bf16(replaced by inline chunked reads). Kept for potential test use.