Repository navigation
Qwen3.5/3.8 on CUDA: kept prompt states in host memory on #836's engine (#726's reuse; on #885) - #888
Open
BobClawblaw wants to merge 49 commits into
Open
BobClawblaw wants to merge 49 commits into
BobClawblaw wants to merge 49 commits into
Conversation
- zig/src/platform: `sys`, imported by every module (zig/build/sys.zig wires it). On POSIX it is std.c, std.posix and std.DynLib unchanged; on Windows win_c.zig gives libc's surface over the UCRT and Win32 (integer descriptors, directory descriptors, readdir, pread/pwrite, high-resolution sleep, monotonic clock, /tmp and /proc/self/exe mapped), win_posix.zig the server's sockets over Winsock, console events as SIGINT/SIGTERM, and LoadLibraryW for DynLib. - std.c and std.DynLib calls go through `platform` (mechanical). - build.zig routes Windows targets to the CUDA build. - Server: socket closes and writes are explicit, the key file's mode check is skipped on NTFS (ACLs), descriptor limits are a no-op. - shard_edit: the shard's map is released before the rename, since Windows refuses to replace a mapped file. - Tests that need symlink privilege or mode bits skip on Windows. zig build test-cuda-host passes on x86_64-windows.
- Kernel fatbins build with nvcc and cl.exe: MSVC's conforming preprocessor is passed on Windows hosts (CUDA 13's headers need it), and explicit kernel instantiations carry the __restrict__ qualifiers their templates have (MSVC counts them in the function type; the generated code is unchanged). - Nemotron's checkpoint source keeps io_uring to Linux; elsewhere its existing thread reads run. - pull_exlr8 raises no descriptor limit on Windows. - modelopt test: the lane's slice partials are allocated. GPUs without clusters (before sm_90) reduce split K through them; the test passed none, so it wrote through a null pointer on Ampere and Ada. zig build native and install build with -Dnvcc=... -Dsm=89, and tf-cuda-test's synthetic suites pass on an RTX 4090.
- pull: where Windows refuses the snapshot's symlink, the snapshot path
becomes a hard link to its blob (CreateHardLinkW through `sys.c.link`):
no extra space, and learning already refuses linked shards. Link
failures are logged with their reason.
- HOME falls back to USERPROFILE (Windows sets no HOME) for the hub cache
and every libc getenv("HOME").
- Tests read optional fixture variables through `platform.getenv`;
std.process.Environ.getPosix does not build on Windows.
- The learned-state symlink tests skip on Windows (symlinks need
privilege); `sys.c.symlink` asks for an unprivileged one.
zig build test passes on x86_64-windows.
- docs/WINDOWS.md: requirements, build, run, how the port works and how it differs from Linux; the README points to it. - -Dstrip reaches tensorfold-native before its install step is made, so the step no longer expects a .pdb the stripped build does not write. - zig fmt on the platform files.
8080 and 8000 are taken by too many other servers; serve listens on 8888 unless --port says otherwise, and the docs and bench scripts follow.
The MLX affine 4-bit checkpoint on the native CUDA server, beside Nemotron: qwen35.cu holds the Metal engine's glue at its bf16 rounding points (gated delta rule, causal convolution, gated attention), cuda.qlinear's affine-4 kernels do the projections, sample.cu draws on the device. Windows of up to 16 rows record each row's recurrence and convolution tail, so the lane core shares rounds across streams and keeps partial copies; a window's rows equal one-row steps bit for bit. draw.zig is the host draw both backends now share. On an RTX 4090: 210 tok/s for one stream, 1250 tok/s across eight; logits agree with an fp32 reference at every position of a 36-token check.
…mory The 2B's CUDA engine now reads its widths from config.json (cuda_geometry.zig): linear value heads over fewer key heads, more query heads a key/value head, an untied head, an A_log stored as bf16. The 27B's logits agree with an fp32 reference at every position of a 36-token check, and its windows equal one-row steps bit for bit; the 2B's output is unchanged. Two Windows driver costs shaped the rest. Each kernel launch grew with the process's allocation count (1.2 ms a launch over 1,900 weight buffers), so a model's weights and scratch now live in one arena (cuda_weights.zig packs each projection through a reused staging buffer: the 27B loads in 25 s, not 200). And a process's first forwards through a buffer stall for seconds, so the engine warms every request path at load and the backend keeps a pool of stream caches that `reserve` allocates and runs a row through once the server knows its stream count. On an RTX 4090: the 27B decodes at 37 tok/s in the runner, 31 to 35 tok/s served, 210 tok/s across eight streams; the 2B stays at 210 and 1,250.
A forward's launches (the embedding to the kept rows' logits) are captured the first time a shape (rows, windows, recording) is seen and replayed after; the attention chunk count moves into the per-forward table so a graph stays valid as caches grow. Prompt chunks stay uncaptured: their GEMMs dwarf the launches, and each length would be a graph of its own. A new graph's first run pays the Windows driver's stall (about a second on a wide model), so `reserve` also captures the shared rounds' shapes, a row of each of 1 to n streams, at load. Bit for bit the same logits; on an RTX 4090 the 2B steps in 2.8 ms instead of 5.3 (served: 324 tok/s, 1,560 across eight streams) and the 27B in 23.8 ms instead of 26.9 (served: 40 tok/s, 243 across eight).
--drafter loads z-lab's DFlash2 draft model behind the lane core's Drafted wrapper: the target writes each row's residual after the drafter's five tapped layers into a ring a stream, the drafter folds those rows into a context of keys and values, and after every verify it drafts a block of seven tokens past the pending one (the target's own embedding and head at its ends, a codebook selector tracing one path through each row's best sixteen). Its projections are quantized to affine-4 on the host at load and run through the target's kernels; its glue is qwen35_draft.cu. The Windows driver stalls a buffer's first kernel and a new graph's first run for seconds, so the drafter pools its contexts and a captured forward now serves every verify shape: its launches cover a bucket of rows and every window, and the kernels read the true counts from the device. On an RTX 4090 the 27B serves 52 to 58 tok/s greedy with drafts against 40 plain, about 100 tok/s on long replies, the same tokens either way; plain decoding keeps its 24 ms step. The 2B is unchanged.
A verify window's rows may form a tree (each row's parent given): the convolution taps, the delta rule and attention follow each row's own branch, and the DFlash2 drafter fills the lane core's planned lanes (each lane its parent's rank-r choice by the codebook selector), so a lone stream verifies up to 12 lanes a round and the planner learns which shapes land. A verify no longer snapshots every row's state (1.3 GB shared, which capped the 27B at 8 rows): each stream logs its rows' decay, write and key, and keep replays the kept path onto the recurrence in one pass (and moves a tree's kept keys down), so any window width fits. One-row steps and prompts update in place as before. The lane host times the target's windows at load, which the depth rule had lacked (it had been capping chains at 3 drafts). The runner's new tree mode verifies tree windows with junk branches against step logits (bit-exact on both models).
What the port serves beyond Nemotron: Qwen3.5-2B and Qwen3.8-27B with its DFlash2 drafter, the speeds measured on an RTX 4090, the load time and why, the timing and chain-only switches, and what is not there.
The lean rule (tools/zig/lean_check.py) counts a comment that runs on to a second line; the Windows port's new files had a hundred such runs. Each is one line of at most 120 columns now, and the checker reports the same 539 problems as main.
direct_io opened files with O_DIRECT on Linux alone. On Windows the same open now asks for FILE_FLAG_NO_BUFFERING (which win_c maps O_DIRECT to), and the 4 KiB-aligned spans it already reads satisfy the sector alignment that needs; a file system that refuses it still falls back to buffered reads. The direct-read test runs on Windows too, checking the executable's bytes against buffered reads of it.
SIG.IGN is a function pointer made from 1, as std.posix has it. aarch64 requires function pointers to be aligned unless the type says align(1), so the cross-build for aarch64-windows refused it. With align(1), as std's own handler type has, tensorfold-native cross-compiles for aarch64-windows with the GB10 (sm_121) fatbins; no hardware has run it.
The Windows page opens with what the build is, what it is tested on and that it cross-compiles for aarch64-windows; the README's model table and the CUDA paragraph say Windows where they said unofficial port.
The no-GPU workflow runs zig build test on windows-latest beside ubuntu-latest, so the platform layer is built and tested on every pull request.
Every other module calls libc through `sys`; the drafter's four getenv calls were the last direct std.c calls in a file the Windows build compiles.
After a prompt the lane core hands the drafter every prompt row's features; the CUDA backend refused any request wider than its tap ring (2,176 rows), so a 27B request with --drafter and a prompt past that ended with RowsNotHeld, where the same prompt without the drafter ran. The ring keeps the newest tap_rows rows, which hold the drafter's whole context window, and absorb already reads only that window of what it is handed; the backend now refuses only rows not yet written or all gone. On an RTX 4090 the 27B with its drafter takes a 6,041-token prompt in 7.4 s, and its reply equals the one with "draft": false.
… attention cuts chunks at logical positions - rowanc held a window's rows as bits of a u32, but prompt chunks run up to 128 rows: rows from 32 on rotated at the wrong positions and attended the wrong keys. A chain row (no parents) now carries its index in its window under the top bit; tree windows (at most 16 rows) keep their bits. - tf_q35_attn_partial cut its 128-key chunks at a key's cache row, so a kept tree path whose window crossed a chunk boundary summed in other chunks than one-row steps and changed bits. Chunks are now cut at the key's logical position (the stream's length plus the ancestors before it). Checked on an RTX 5090 (Linux, sm_120), tf-qwen35-cuda over 300 ids on Qwen3.5-2B and Qwen3.8-27B: step bits unchanged, window and tree equal step byte for byte (tree differed from row 126 on the 2B and 254 on the 27B before). Prompt mode's last row against step's at 16 to 300 ids: KL at most 0.003 and the same argmax in 20 of 20 (before: 27B at 40 ids KL 2.3 with a different argmax, at 128 ids KL 1.35).
zig build test-golden and zig build fuzz-cortheon exist on every non-macOS host since the golden step moved into zig/build/golden.zig, so the golden check compiles against the platform layer on Windows: - the platform's posix gains sockaddr.in, timeval, connect and kill for a child's handle (TerminateProcess; HUP has no equivalent), setsockopt turns a timeval for SO_RCVTIMEO or SO_SNDTIMEO into Winsock's milliseconds, and a socket read that times out is WouldBlock, as std.posix reports it. - the check carries SystemRoot into the spawned server's environment, which Winsock needs to start, and drains the server's stdout after the port line on every host: the fake server logs each request there, and a full pipe (4 KiB on Windows) blocked it after sixteen cases. - on Windows the check cannot send SIGTERM or SIGHUP, so the sigterm-exit-0 outcome and the cases after a key rewrite are skipped and printed as such. On Windows the goldens give 306 of 315 equal with the same nine known differences as Linux, and the fuzz step passes its eight tests; the no-GPU workflow's Windows job runs both.
The loader tried only libcublasLt.so.13 and libcublasLt.so, so the cublaslt suite of tf-cuda-test reported LibraryUnavailable on Windows; there it opens cublasLt64_13.dll from the toolkit's bin folder.
The platform handed O_APPEND to the C runtime as _O_APPEND, whose write seeks to the end and then writes, so two threads appending at once could overwrite each other's bytes: the request log lost lines under concurrent requests, which the server goldens' request-log case caught on a slow runner. A descriptor opened with O_APPEND now writes through WriteFile at the end-of-file offset, which the kernel places whole, as POSIX does; the file keeps its write access, so ftruncate on such a descriptor still works (prompt_imprint repairs a torn index that way). A request_log test appends 200 lines from each of 8 threads and expects every line whole; it lost lines on Windows before this change.
The lifecycle/request-log outcome said only that the lines differ; it now gives both logs' line counts and the first line that differs, cut to 100 bytes each, so a mismatch seen on a CI runner can be read from its log.
…wer: batched prompt passes (prefill_many, addStreams), a backend's refusals for a memory guard, follow() for a tensor-parallel rank 1, shared-pool admission fields, image and video inputs for the native server (vision_inputs, media_fetch, qwen_image, stb_image)
A checkout with core.autocrlf, as GitHub's Windows runners make, gives the frozen request log CRLF line endings while the server writes LF, so the lifecycle/request-log case differed at its first line's end on windows-latest and nowhere else. Both logs compare with CRLF read as LF; the other goldens are parsed as JSON and never saw the endings.
…cel; lean rule and zig fmt setup() returned cancel()'s true for a background warm pass that yields to a newer request, so the caller launched the job cancel() had just finished (SEGV in lane_host_warm_test). It now returns false, as for a drop. Comments one line each; lean_check skips zig/vendor/ (stb_image.h, as on ashhart#726). lane_host.zig and lanes/engine.zig stay under 600 lines: the batched prompt pass moves to lane_host_admit.zig (many, one) and lanes/open.zig (addStreams, adoptPrefilled), aliased where they were.
…rozen words (the errors/content-image golden)
…nsorfold-native takes memcpy from libSystem again (zig build test on macOS)
…eports; images wait for the shared decoder; no refused hook - prompt_cache: Entry.pins, Store.pin/unpin. victimBut passes pinned entries over, so a later job's lookup in admit.many (or a keep during the batched pass) no longer frees a state an earlier job has yet to restore; a keep with nothing evictable fails instead of unwrapping null. LaneHost.setup pins, reported and requeue unpin. - vision: zig/vendor/stb dropped and no longer compiled. qwen_image.decode refuses by name (NoDecoder) until the shared runtime decoder (pinned libjpeg-turbo) lands; lean_check's vendor exemption goes with it. - lanes: Backend.VTable.refused, Engine.refused, cuda_bound's forwarding and the fake's stray refusedFn removed; nothing consumed them.
…'s widest stream The chunk width came from the longest stream or window in the forward, so a 201-key row cut its keys into 128-key chunks alone and 256-key chunks beside a stream past 8192, and a window crossing 8192 recut its earlier rows; either changed the fp32 sums. Each row now takes chunk_keys(its logical keys), the width its one-row step uses, in both the partial and the merge kernels, and the partials sit at a fixed max_chunks stride. tf-qwen35-cuda takes TF_Q35_PREFILL=N (start after N cycled ids) and TF_Q35_BESIDE=N (step an N-token stream in each forward), which cover both cases: before this change, windows from 8151 keys kept 36 of 120 rows and trees 35, and a 151-key stream beside one from 8189 kept 4 to 11 of 120; after it all of them equal one-row steps on the 2B and 27B.
…fits its selector rank The integer reader accepted zero, so a conv_group_size or num_key_value_heads of 0 reached a remainder at admission; sizes and divisors are now positive (mask_token_id alone may be 0). The host staging area held 16 rows at rank 256, so a larger selector_rank overran it; it is now sized for the admitted rank. Config.parse takes the bytes, so a host test covers both.
…eed only at its last unpin; a job reports resumed before it unpins
# Conflicts: # build.zig # zig/build/cuda.zig # zig/src/native/cuda.zig
A prompt chunk can carry an image prompt's rows (Part.images): the tower's features replace its placeholder rows after the embedding (tf_q35_splice), and tf_q35_rope turns each row by its t, h and w positions (interleaved sections 11, 11, 10) from a staged table. Rows after an image prompt rotate at their position plus the stream's rope_delta (Seg.delta), so decode, verify windows and trees keep their captured graphs. Text rows rotate at one position on all three axes, the same arithmetic as before: their logits keep their bits.
The checkpoint's Qwen3_5VisionModel in bf16 (ashhart#726's tower and kernels, its Linears on cuBLASLt): width, heads, depth and MLP read from vision_config, so the 2B's tower (24 blocks of 1024, heads of 64) runs beside the 27B's and the 35B-A3B's. Its weights and one 16,384-patch call's scratch are one allocation made at startup and counted as model memory. At prefill a request's images go through the tower; their rows and get_rope_index's positions feed the prompt chunks, and the stream keeps rope_delta. Video parts are refused by name.
vision: each 0.6.6 fixture's patches through the tower, rows' cosine to 0.6.6's (tools/qwen_vision_fixtures.py). mrope: a prompt through the image table with text positions must equal prompt's logits bit for bit; TF_Q35_SHIFT=k moves the h axis, a control whose logits must differ.
The ashhart#853 merge removed learnBytes (ashhart#836 reads learn_bytes in place); its doc line stayed above toGib.
…#726 Two image prompts with the same text render the same ids, so keyed by tokens alone one could resume the other's state. Request.cache_key holds the prompt with each image's placeholder rows as ids from its content; lookups, keeps and resumes take cacheKey() (the prompt when empty). The test is ashhart#726's.
…m each image's content image_key.cacheKey: placeholder run k (Qwen's image or video pads, GLM's image token) takes ids hashed from image k's grid and patches and the row's place in its run; text keys as itself, so a change to a later image keeps the prefix before it. Runs that do not pair with the images key by all of them.
…d a trailing blank line - vision_inputs opens libtfvideo through sys.DynLib (LoadLibraryW on Windows, as tfvideo.dll): std.DynLib has no Windows. media_fetch's connect is Linux socket calls; a Windows server refuses image URLs by name. ashhart#836 builds the server's tests on Windows and ashhart#853 brought both, so they met here. - tf-qwen35-cuda's header and two comments within one line and 120 columns. - docs/vision.md ends in one newline (git diff --check).
…ry (ashhart#726's reuse) cuda_reuse.zig is the family's Snapshots target: a save copies a stream's recurrences and conv tails whole, each attention layer's key and value rows up to the mark, and its rope_delta, into pinned host memory; restore copies them back. The prompt pass restores a kept state and calls the lane host at the store's marks only where a cold pass cuts a chunk (every max_rows rows from 0, or a request's chunk start), so every resumed reply keeps a cold pass's bits. The native host takes any family's snap_* functions; a snap_host family's cache is --prompt-cache-gib of host memory (8 GiB unasked, at most half of MemAvailable), on planned rules over its prompt grid.
BobClawblaw
force-pushed
the
qwen35-cache
branch
from
October 15, 2026 15:14
afb95fb to
b0aab1f
Compare
8 tasks done
Owner
|
Thank you, @BobClawblaw, for preserving cold-prompt chunk cuts and holding host copies through restore. At b0aab1f, Please send a follow-up that guards piece construction, verifies full image identity and handles the cache format before restoring any saved state. Posted by Ash's AI Assistant Synthena |
Each image's rows must reach a mean cosine of 0.999 and a worst row of 0.94 against 0.6.6's bf16 rows (both settable). A fixture with 0.6.6's fp32 rows (qwen_vision_fixtures.py --fp32) judges the tower against those instead: it may not sit further from them than 0.6.6's own bf16 tower, beyond 0.001 on the mean. The 2B's tower is fragile in bf16 (0.6.6 against itself drops to 0.45 on a row), so only the reference can bound it.
… from it lanes.Image.digest covers the grid and every patch bit. image_key seeds each run's ids from it and digests each image once. The 31-bit ids only find a kept state: the backend compares whole digests before it restores one.
…s cache format, its images cuda_reuse.zig's pieces are computed by index from the geometry, so no layer or head count passes a fixed array (33 attention layers of 32 key heads did). A kept state records its cache format, and its byte layout comes from it: bf16 rows, or with an engine that has --kv-cache fp8, e4m3 rows and their scales as qwen35_kv8.cu lays them out. It also records the digest of every image whose placeholders start before its position. A restore of another format, length or image set is refused before the stream's cache is touched, and the prompt is prefilled from the start.
…ift (TF_Q35_DELTA) TF_Q35_RESUME=1 restores the stream after its prefill into a fresh, poisoned cache through cuda_reuse.zig's save and restore before step, window or tree; TF_Q35_RESUME=K does it at position K of a prompt or mrope pass. TF_Q35_DELTA=d rotates the rows after the prefill at position + d, as an image prompt's reply. cuda_reuse.zig's saveCache and dropCopy take a cache.
# Conflicts: # zig/tests/qwen35_cuda_main.zig
Contributor
Author
|
Follow-up at f8a9973: commits 6e29ebe, e64191a and 78a222b, then #885's head merged in.
Checks, RTX 5090:
|
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
2 participants
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.
Kept prompt states on #836's Qwen engine
Asked for on #726: its host-memory prompt reuse, brought onto #836's engine as the lane core's kept-prompt store, with image, resume and concurrency checks. Stacked on #885 (image input on #836 and #853), because image prompts are part of what the cache has to get right. Review the last three commits:
lane host: the prompt cache keys a request by cache_key(from Qwen3.8-27B on CUDA: image input and --kv-cache fp8 (256K with vision on a 5090) #726): two image prompts with the same text render the same ids.Request.cache_keyholds the prompt with each image's placeholder rows as ids drawn from its content, and lookups, keeps and resumes usecacheKey(). Qwen3.8-27B on CUDA: image input and --kv-cache fp8 (256K with vision on a 5090) #726's lane-host test comes with it.server: an image request's cache key:image_key.cacheKey. Placeholder run k (Qwen's image and video pads, GLM's image token) takes ids hashed from image k's grid and patches. Text keys as itself, so a change to a later image keeps the prefix before it.qwen3_5: kept prompt states for the shared cache, in pinned host memory:cuda_reuse.zigis the family'sSnapshotstarget. A save copies a stream's recurrences and conv tails whole, each attention layer's key and value rows up to the mark (one copy per key head), and itsrope_delta, into pinned host memory. Restore copies them back.snap_*. Asnap_hostfamily's cache is--prompt-cache-gibof host memory (8 GiB unasked, at most half of MemAvailable), on planned rules over its prompt grid. This is Qwen3.8-27B on CUDA: image input and --kv-cache fp8 (256K with vision on a 5090) #726's wiring.Images and the cache. #853's server sets
history_lento 0 for image prompts ("image prompts keep no prefix for a later turn, as Python"), so the store plans no marks for them and they neither keep nor resume. This PR leaves that policy alone. The content key is what makes it safe to turn on later.Checks (RTX 5090, Qwen3.8-27B + DFlash2,
--parallel 2 --context 65536 --vision)#726's
exact_cache.py, cache on against cache off on this same build: greedy 256-token replies, each asked twice, over 1.5k to 50k-token prompts, a repeated prompt and a 3-turn chat. 20 of 20 replies are byte-identical to the cache-off run, and every second ask resumed:Images and concurrency (prompts of about 5–7k tokens, against cold replies recorded on a cache-off server): 12 of 12 replies equal to cold:
Host:
zig build testandtest-cuda-hostpass, with new tests for the cut-point rule, the state's pieces, the image key, and #726's lane-host key test.zig fmtis clean, and the lean check reports 0 new problems.Not run: Windows (host memory is sized from
/proc/meminfo, so a Windows server gets no cache yet), a GB10, and the 35B-A3B, whose MoE entry on #856 needs the samesnap_*exports once this lands.