Repository navigation
Conversation
Step: test-qwen-serving-host. Old pins: e166f3f. New base: 063d881 (origin/main). Sets the five kernel hashes (experts.cu, experts_pack.cu, experts_prefill.cu, qmm_group.cu, qmm_prefill.cu), the qmm_group prefix and the launch-sequence hash to 063d881's bytes. The prefix is now group[0..16095]: line 1 lost 37 bytes and the rest is byte-equal, so it covers the same text e166f3f's group[0..16132] did. No code changes; 2/2 tests pass. Recorded, as merged and reviewed: - #2229 (3d4f6aa): a prompt chunk's last layer runs q, attention, gate, o and the MLP over its last 128 rows (the launch-sequence pin). - efdf28f, fcb1529, f3bcb78, ecef17d, 2ebc3c2, 42d3461 (lean, no PR number in the history): line-1 comments of the five kernel files only.
Step: test-gradient-serving. Old pins: e166f3f. New base: 063d881 (origin/main). Sets the hash of the 46 changed entries in gradient-serving.tsv to 063d881's bytes; paths, modes (all raw) and note columns are unchanged. pocket-serving-approved.tsv needs no edit: its entries for these files have prefix 0, so they never read the manifest hash. No code changes; 1/1 test passes. Recorded, as merged and reviewed: - #1784 (cf63ff6): Flash Next at --tp 2 splits dense output columns with exact FP32 gathers. - #1784 (cde4fe7): the BF16 prompt tile folds a pinned (cut) weight's whole K-slice order. - #1942 (15f118a): a shared round's DeltaNet steps run as one launch over its lanes. - #1792 (52b9f90): the whole shared expert's down writes straight into its output slot. - #2160 (a54748a): beside another open stream a greedy chain stops at 0.6 a draft; wide rounds draft 6 deep. - #2246 (4294752): at --parallel 1 on one GPU a prompt pass takes 8,192 rows. - #2214 (e4d7f13): kept prompt states with drafts across ranks. - #2230 (1f2c993): Gemma 4's PLE kernel in its own image; controls and blob constant. - #2103 (a3962b4, 194115d): synchronous downloads stage through a pinned bounce, on integrated devices only, and the bounce counts in memory use. - #2289 (e9ee738): the remaining synchronous readbacks use that staging; close() releases it. - #2127 (b2c277a): the lane host's KV ledger counts the paged pool's promises. - 4fb2f65, 60d80da, 8cc6ac4, 36a87df, 41ff25d, efdf28f, fcb1529, f3bcb78, ecef17d, 2ebc3c2, 42d3461, a86c31c, 1913a46, 80218cd: comments, verbatim moves, test-block imports, exports, and opt-in or off-by-default paths only.
files: prefill_attention.cu (new constexpr P256 and IDS flags, false in every existing kernel), prefill_attention_rows.cuh (the 256-position page items and their page lookup) and cuda/abi.zig (memory_type_host). Its review and device runs were done before it landed, so the pins take main's bytes.
…ER and swapped red and blue The render flags were 0x01 | 0x10, commented as FPDF_ANNOT | FPDF_PRINTING, but in PDFium 0x10 is FPDF_REVERSE_BYTE_ORDER (FPDF_PRINTING is 0x800). PDFium then wrote RGBA, which the JPEG encoder reads as BGRA, so every rendered page reached the vision tower with red and blue swapped (#2386). The flags are one named constant for both render calls, pinned by a test.
…t its context The rows store's write worker had a fixed 1 GiB permit, and a save whose new bytes passed it was skipped with StagingFull. At about 65 KB a token, a 27B conversation whose first save or post-compaction prompt held more than about 13K new tokens was never written, nor any later save of it. rows.Options.longest sizes the permit to one state of that many tokens with no rows on disk (coldCharge), never under the 1 GiB floor nor past the RAM store's budget; the 27B host passes its --context. Other hosts keep 1 GiB.
zig/build/cuda.zig was at the 600-line limit, so every new CUDA hook had to move lines out first. Two contiguous groups of host-test registrations from hostTests move verbatim, each behind one pub fn called once from where the group stood, so steps and dependencies register in the same order: - cuda_contract_host.zig steps(): the CPU contract references (Flash Next writeback, utility, row-best and QSA; cuda_row_contracts and cuda_window_host hooks; FP64 quant and projection fixtures). - cuda_server_host.zig steps(): the HTTP server's host tests over server_tests, through test-server-cpu and the off-macOS goldens and fuzz step; it returns template and server_tests for the tests after it. All 134 moved lines and the golden.zig import are byte-identical; the new parameters keep the old variable names. cuda.zig: 600 -> 467 lines. zig build -l is unchanged (empty diff) by default, with -Dhip=true and for x86_64/aarch64 Linux and x86_64 Windows targets.
…n, as step5-filter is Linux's
Owner
Author
|
Stopped at #2388. |
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
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.
CI for a landing train: #2418 #2387 #2388 #2420 #2163 stacked on main in that order. Nothing merges from this PR: each PR above lands on its own once this tip is green, and this PR is then closed.