GLM-OCR on HIP: the publisher's BF16 checkpoint, exact against its six model-card tasks by laudney · Pull Request #1940 · ashhart/TensorFold

laudney · 2026-11-02T14:23:44Z

Part of #1927.

This serves the publisher's BF16 zai-org/GLM-OCR (revision 2e85a62840ccac27daa451df36c736c4636b8628) natively on HIP. The CogViT image encoder and the GLM decoder run in Zig and HIP, with no Python or libtorch in the served engine. It is the checkpoint other engines publish numbers for and the one users download first, so serving the publisher's weights matters for benchmarking and adoption. Every reply is checked against the model card's own Transformers example.

	Before (main)	After (this PR)
`zai-org/GLM-OCR` (BF16) on HIP	not supported	served natively on Radeon AI PRO R9700 (gfx1201)
Python or libtorch in the served engine	—	none; one pinned Transformers 5.5.4 oracle under `tools/zig/` for checks
Six model-card tasks vs HF (recognition, description, text, formula, table, extraction)	—	exact reference IDs and text over HTTP, plain and copy-drafted
Concurrent pairs vs solo	—	identical IDs
Batched verification rows (widths 1-32) vs single-row projections	—	23,519,232 BF16 values, all exact
Strict structured extraction	—	every schema field, with drafts off and on, identical IDs
Kernel contract on gfx1201	52 kernels pass gates 1-4	70 kernels pass gates 1-4, 5/5 controls red
Inherited limit problems (live gfx1201 / offline, 4 chips)	124 / 503	124 / 503, the same entries (0 new)
Existing models	—	unchanged: new image budgets, runtime libraries and cache rules are opt-in

Prerequisites #1929 (prompt-cache rules), #1930 (rocBLAS projections) and #1933 (no-reasoning replies) are merged. This branch still carries their earlier copies at its base, so a rebase onto main leaves only the GLM-OCR commits. The GLM-OCR review is the last ten commits, 4e64877cb^..a8594af84.

Commits

Checkpoint and image preparation: config, the 526-tensor BF16 schema with refusals for missing, extra or wrongly typed tensors, the image processor, and M-RoPE positions, with processor fixtures.
BF16 weights on HIP and matrix products: checked weight ownership and the BF16 projections on HIP: rocBLAS BF16 projections with FP32 sums #1930's rocBLAS module.
The pinned Transformers oracle, tools/zig/glm_ocr_reference.py. It has four commands:
- capture runs the model card's example and saves the reference evidence.
- processor writes the committed fixtures.
- attention writes the SDPA fixtures for the AOTriton ABI test.
- examples rebuilds the six model-card tasks.
feat(picture): explicit source image budgets. picture.Limits and decodeWithLimits let a caller admit a full linked image before resizing. The default limits are unchanged, and the existing callers keep decode.
The image encoder:
- CogViT with MIOpen convolutions (hip/convolution.zig), restricted to one GEMM solver family, so its arithmetic does not change between searches.
- AOTriton flash attention (hip/attention_aotriton.zig).
- The pointwise kernels in kernels/hip/glm_ocr/pointwise.hip.
Image and text decoding on HIP, with the fixed-algorithm hipBLASLt decode path (hip/blas_fixed.zig).
refactor(hip): share native open and image offers between Qwen3.5 and GLM-OCR. Qwen3.5 behaviour does not change.
Batched exact verification: the vocabulary rows of a window are projected on the GPU with the single-row algorithm.
Serving exact image replies: registration and image offers, and pub const reasoning = false; (Server: one engine flag for models whose replies carry no reasoning block #1933), because GLM-OCR writes no reasoning block. Its prompt still follows enable_thinking.
Prompt-state reuse with exact arithmetic: image repeats are keyed through Rules.repeat_keyed (Prompt cache: an opt-in rule for families to count keyed image prompts as repeats #1929). This commit also adds the recipe docs/recipes/glm-ocr-hip.md.

Runtime libraries (optional, loaded at run time)

Library	Used for	Notes
rocBLAS (#1930)	BF16 projections
hipBLASLt	fixed decode algorithm	`libhipblaslt.so.1`
MIOpen	image convolutions	keep its find database on persistent storage
AOTriton 0.11.1	flash attention	MIT; `LICENSES/AOTriton-MIT.txt` and `THIRD_PARTY_NOTICES.md` are added; `TENSORFOLD_AOTRITON` selects it
libjpeg-turbo 3.1.4.1	JPEG decode equal to the reference	the repository's pinned decoder

Known limits

Prompt-state reuse restores a prompt of the same length. The engine keeps full-prompt arithmetic for fidelity, so a longer next turn prefills again rather than resuming at an arbitrary cut. Native OCR models on HIP: dots.mocr, GLM-OCR, LightOnOCR-3 and Nemotron Parse, CUDA to follow (tracker) #1927 tracks a shared rule for exact-length caches.
These are not qualified: the checkpoint's separate MTP head, tensor parallelism, fp8 KV, Sliding Weights, YaRN, CUDA and Metal.
build.zig gains one line after const test_step. core: a GGUF reader and an exact repack of GGUF's quantized blocks (#1920, steps 1-2) #1926 and dots.mocr on HIP: the publisher's BF16 OCR checkpoint, token-exact against its model card #1938 add a line at the same place, so whichever lands later keeps both.

Gates

Kernel contract on a Radeon AI PRO R9700 (gfx1201): SUMMARY gfx1201 70 kernels: 70 pass gates 1-4, 0 fail, 2 known findings; 5 controls: 5 red as planted, 0 missed; 124 limit problems.
- The upstream/main base reports 52 kernels ... 124 limit problems on the same GPU, with the same limit entries.
- All 48 GLM-OCR record combinations of kernel and chip match the code-object metadata.
Host: test-hip-blas-host, test-glm-ocr-host, test-glm-ocr-hip-host, test-glm-ocr-native-host and test-lane-api-host pass. zig build test: 470/470 steps. zig build test -Dhip=true: 502/502 steps.
Commits: every commit builds native-hip for gfx1201.
Lean checker: zero new problems over the base.

ashhart · 2026-11-02T17:58:41Z

Thanks @laudney. We reviewed this independently, and it lands. The shared changes keep today's defaults: Info.reasoning true, the picture.Limits defaults, repeat_keyed false, hip_serve.Options.vision null, and the new refusals only for families that declare them. The host suites pass under -Dhip=true (485 passed, 1 skipped). We have no AMD GPU, so the kernel contract, the gfx1201 reference ids and the six model-card cases are your receipts. It's gated and queued, and we've written the changelog line.

laudney · 2026-11-02T18:31:19Z

The #1938 review asked dots.mocr for pinned sources, a device receipt and a per-family reasoning assertion. This PR gets the same now, in two new commits on top (e0546d811, 1aab5dd64), with no force-push.

Item	Before	Now
Publisher sources	the oracle checked the checkpoint revision only	the SHA-256 of the 8 non-weight sources at `2e85a6284` (configuration, generation and processor configs, tokenizer, chat template, model card) and of the 5 Transformers 5.5.4 files that hold GLM-OCR's model and processor code, refused on any difference; `python3 tools/zig/glm_ocr_reference.py sources <checkpoint>` checks them without a GPU or a model load
Reasoning assertion	the registry test from #1933 names no family	the GLM-OCR flag test asserts `!reasoning(glm_ocr.native)`
Device receipt	the PR body's account	below, from a new build of the PR head
`capture` receipt	revision and runtime	also the pinned source hashes and runtime versions

GLM-OCR's model code is not in the checkpoint: Transformers 5.5.4 ships it. So the oracle pins those library files by path inside the installed package, which it finds without importing Transformers.

Pinned sources (SHA-256)

File	SHA-256
`README.md`	`172334449c3df98650bc9a42781c254c64c9d466b440e2ca8ce31f99829212b2`
`chat_template.jinja`	`062e7ee4cc8defa88a5938b5d456dc60366ffd80647f918946ee747bf09ddc7c`
`config.json`	`ad537817b2becde827cbb23569a1dd83a352ee64cc34ce751a4ef000ecce4b5f`
`generation_config.json`	`4e2a3412f65b2a21d315d928986c55723d3e60dfb92c3982a5b9fd56835b0aa5`
`preprocessor_config.json`	`02cc50c36240882ae35e8cd4077a25a379664108185d728d261cb785aefeccff`
`processor_config.json`	`a56fbd559186fefd6c22a3acc6a393a18a9fe52c264692a122a143d24042112b`
`tokenizer.json`	`aa0fd058c73a5718bb191f6672dc16d122ee0147b20c123d1726514298f9968a`
`tokenizer_config.json`	`95abd49942be9cdc6e2653528aed7b4c1cb9e180828421d3488ab38da3057016`
`transformers/models/glm_ocr/configuration_glm_ocr.py`	`7015c0aa18811ab058e228a9c343bc07bf266abb2753a66a2c371720b6612c53`
`transformers/models/glm_ocr/modeling_glm_ocr.py`	`f667126388eacde9ca93983c1fc34488f4c169717a6ff5900c1479d2c6bb7db4`
`transformers/models/glm46v/image_processing_glm46v.py`	`41fa88e258a4b291f167d77085eee989dbebff1776c336121e2f4185b6e0eef1`
`transformers/models/glm46v/image_processing_pil_glm46v.py`	`3506c3f76284c90d3ed9d4e44e5091cb09ceb4fcd8f382cb2a90aa18923c7598`
`transformers/models/glm46v/processing_glm46v.py`	`91ba14d315a2059e4f0f811618394f201401a9858648c1f66c4637f25fe628db`

Device receipt

One Radeon AI PRO R9700 (gfx1201), selected by ROCR_VISIBLE_DEVICES and checked by PCI address (0000:c3:00.0); other cards ran unrelated work. HF ids come from the existing Transformers 5.5.4 oracle runs; the oracle was not re-run. A = sha256 of comma-joined ASCII ids (the server's token_sha256). R = sha256 of little-endian u32 ids. Only 16 hex digits are shown; we can attach the full values on request.

identity	value
candidate	`1aab5dd648a6` (PR head); `tensorfold-native` 1.2.1 sha256 `1c09d827c9cef97cd55dd165e183a516be8a77e94bca9de32435ce98152e5a93`
base	`71932143f95a` (upstream/main); `tensorfold-native` 1.2.1 sha256 `2e17004a34f084c3e2d1f84557cdb0c5f2ede94ff806913e94acf974b9806e33`
build (both)	`zig build native-hip -Dhip=true -Dhip-arch=gfx1201 -Dhipcc=/opt/rocm/core/bin/hipcc -Dhip-device-lib=/opt/rocm/core/lib/llvm/amdgcn/bitcode -Doptimize=ReleaseFast -j8`
toolchain	Zig 0.17.0; ROCm 10.1.0 (rocm-gfx120x); hipcc/HIP 7.16.26385; AMD clang 24.0.0git
libraries the server loaded	rocBLAS `librocblas.so.5.7` `75645cbd1f2dd32f`; `libhipblaslt.so.1.4` `3d414626f84ecc43`; MIOpen 3.6.1 `libMIOpen.so.1.0` `0549aa0661734db6`; HIP `libamdhip64.so.7.16.26385` `bddbf3553ee64ae5`; AOTriton 0.11.1 `libaotriton_v2.so.0.11.1` `e0638806efa5d35c` + gfx120x images (1888 files, tree `f01f1ecb70206d4f`); libjpeg-turbo 3.1.4.1 `b6a5f4e6066da95c`; libtfgrammar (xgrammar v0.2.8 `97787376`) `0a8a1c3f27c4eef5`
checkpoint	`zai-org/GLM-OCR` @ `2e85a62840ccac27daa451df36c736c4636b8628`; `model.safetensors` (2,650,579,464 bytes) `a16eb0de98d199293371c560f95f83130d2a2c9612449df16839f08ff9498815` (equal to the HF LFS id)
`glm_ocr_reference.py sources`	pins pass; config `ad537817b2becde8`, tokenizer.json `aa0fd058c73a5718`, chat_template.jinja `062e7ee4cc8defa8`, preprocessor `02cc50c36240882a`; Transformers 5.5.4 modeling_glm_ocr.py `f667126388eacde9`, processing_glm46v.py `91ba14d315a2059e` (config and modeling equal in all 6 HF receipts)
assets	car `8cd28b372908bcaa`, bee `8b21ba78250f852c` (documentation-images @ `9dd8e471`, original bytes); text `32b9c78df373c021`, formula `7a91aa1b95543f76`, table `6b9f3dc2af058e12`, extraction `501a86f14f69aaa7` (fictional fixtures for the model-card prompts)
HF reference	Transformers 5.5.4, torch 2.10.0+git8514f05, HIP 7.2.53211, MIOpen 3.6.1, SDPA, bfloat16, greedy, same GPU
server	`serve <ckpt> --host 127.0.0.1 --port 18897 --name glm-ocr --vision --context 8192 --parallel 2 --prompt-cache-gib 1 --mtp-drafts 0 --max-tokens 512 --temperature 0 --no-update-check --idle-unload 0` (recipe flags); a second process with `--prompt-cache-gib 0` for the cache-off rows; both SIGINT exit 0
requests	model-card prompts, data URLs, `return_token_ids: true`, `temperature 0`, max_tokens 512 (bee 128)

check	native	HF / expected	equal
before: base `serve`, same flags	exit 1: `the native HIP engine has no backend for glm_ocr checkpoints yet; serve with --engine python`	refusal before weight load	n/a
car: prompt 403 ids; output 7 ids, last 59253 (EOS), `stop`	prompt A `148bff80c05084e5`; out A `93aec362c628084a` R `02cc86b94f00562d`	prompt A `148bff80c05084e5`; out A `93aec362c628084a` R `02cc86b94f00562d`	yes
bee: prompt 6156 ids; output 6 ids, last 59253 (EOS), `stop`	prompt A `5f659154f6e7b7da`; out A `c8698f4784ceb1f5` R `be2ba128f355f4e1`	prompt A `5f659154f6e7b7da`; out A `c8698f4784ceb1f5` R `be2ba128f355f4e1`	yes
text: prompt 252 ids; output 26 ids, last 59253 (EOS), `stop`	prompt A `17affe56723c83e3`; out A `db7213f5ace9e43a` R `8eeebd358f0af890`	prompt A `17affe56723c83e3`; out A `db7213f5ace9e43a` R `8eeebd358f0af890`	yes
formula: prompt 223 ids; output 21 ids, last 59253 (EOS), `stop`	prompt A `54954047df4014b4`; out A `cec8bf3afa2b36ee` R `c5a936f8c02793e3`	prompt A `54954047df4014b4`; out A `cec8bf3afa2b36ee` R `c5a936f8c02793e3`	yes
table: prompt 192 ids; output 58 ids, last 59253 (EOS), `stop`	prompt A `f6712bfa91c76445`; out A `b85de97f46b59f52` R `6f2f5b715ffb12cc`	prompt A `f6712bfa91c76445`; out A `b85de97f46b59f52` R `6f2f5b715ffb12cc`	yes
extraction: prompt 850 ids; output 155 ids, last 59253 (EOS), `stop`	prompt A `ed83816e526add68`; out A `f12b92635ade64f0` R `d081e35d955f873b`	prompt A `ed83816e526add68`; out A `f12b92635ade64f0` R `d081e35d955f873b`	yes
all 6, cache on: fresh plain (cached 0)	ids = HF ×6	HF	yes
all 6, cache on: resumed plain / resumed drafted (7/110 accepted); cached car 401, bee 6154, text 250, formula 221, table 190, extraction 848	ids = HF ×12	HF	yes
all 6, cache off: fresh plain / fresh drafted (7/110 accepted)	ids = HF ×12	HF	yes
concurrent pairs table‖extraction, car‖bee, text‖formula: plain / drafted (7/109 accepted); log `shared verify: 2 streams`	ids = HF ×12	HF	yes
seeded sampling text, T 0.7, top-k 32, top-p 0.9, seed 9281: solo, repeat, drafted, concurrent plain, concurrent drafted; 26 ids, `stop`	A `edeac8e219d80693` R `f6d9d06c30c65f41` ×5	equal across 5 cells; differs from greedy	yes
seeded sampling formula, T 0.7, top-k 32, top-p 0.9, seed 9281: solo, repeat, drafted, concurrent plain, concurrent drafted; 21 ids, `stop`	A `cec8bf3afa2b36ee` R `c5a936f8c02793e3` ×5	equal across 5 cells; equals greedy	yes
seeded sampling table, T 0.7, top-k 32, top-p 0.9, seed 9281: solo, repeat, drafted, concurrent plain, concurrent drafted; 58 ids, `stop`	A `b85de97f46b59f52` R `6f2f5b715ffb12cc` ×5	equal across 5 cells; equals greedy	yes
seeded sampling extraction, T 0.7, top-k 32, top-p 0.9, seed 9281: solo, repeat, drafted, concurrent plain, concurrent drafted; 155 ids, `stop`	A `f12b92635ade64f0` R `d081e35d955f873b` ×5	equal across 5 cells; equals greedy	yes
strict `json_schema` extraction (xgrammar): plain / drafted (26/106 accepted), 151 ids	strict JSON, all 11 fields; A `aa6e81dd24c5816d` / `aa6e81dd24c5816d`	fixture fields; plain = drafted	yes
client closes at 0.25 s, before engine submit (body/JPEG decode); then a background car probe and car	no `ended` line, no engine work; probe 0.19 s (before cut 0.19 s); car = HF	dropped, no leaked admission count	yes
client closes in prefill after the stream opens (fresh bee bytes), then car	`client-left`, tokens 0; car = HF	0 tokens, exact next	yes
client closes in decode after 8 answer events, then car	`client-left`, tokens 8; car = HF	exact next	yes
context edge: 403 prompt + 7,789 max_tokens / + 7,790	200 (car = HF) / 400 `context_length_exceeded`	200 / 400	yes
invalid image bytes / truncated JPEG	400 `image 1 is not PNG, JPEG or WebP; send one of those` / 400 `image 1 could not be decoded; send PNG, JPEG or WebP`	400 / 400	yes
`video_url` part, then car	400 `content parts must be text or image_url; audio and video inputs are unsupported`; car = HF	400	yes

In all replies, the OCR text is in message.content, with 0 reasoning tokens. Copy drafts made proposals only for extraction, table; for bee, car, formula, text the drafted = plain result is trivial. Seeded sampling changes only text; the other sampled replies equal greedy, which the peaked OCR distributions explain. The extraction reply keeps the model's Markdown fence, as HF does; only the json_schema request returns strict JSON.

Kernel contract and limits (the two new commits change no kernel)

	`upstream/main`	this PR
gfx1201 contract	52 kernels pass gates 1-4, 5/5 controls red	70 kernels pass gates 1-4, 5/5 controls red
Limit problems (live gfx1201 / offline, 4 chips)	124 / 503	124 / 503, the same entries

Ordering: #1929, #1930 and #1933 are now on main as your versions, so this branch shows conflicts with main in build.zig, zig/build/hip.zig, zig/src/hip/root.zig, zig/src/native/hip.zig, zig/src/platform/optional_libraries.zig and zig/tests/hip/blas_test.zig. They are the copies of those PRs. A rebase onto main leaves only the GLM-OCR commits. We do not force-push open PRs. If you prefer, we can push a merge of main instead.

GLM-OCR support starts with the checks that need no GPU. A new glm_ocr family admits the BF16 checkpoint graph, reads its model and processor settings, and prepares grouped temporal image patches with bounded geometry and rotary coordinates before any device memory is allocated. Host tests compare that preparation with fixtures taken from the Transformers processor, and test-glm-ocr-host and test-glm-ocr-fast run them as part of the default test step.

The native engine needs the checkpoint on the GPU in a form the shared rocBLAS projections can read directly. GLM-OCR now uploads its tensors unchanged into one admitted HIP allocation, finds them by name, and refuses a checkpoint that does not fit before it allocates. Projections reuse hip.blas, so each dot sums in FP32 and rounds to BF16 once, with bias rows added before that rounding. test-glm-ocr-hip-host covers ownership and refusal on the host, and build-glm-ocr-weights-test builds the device upload check.

Native GLM-OCR checks compare against the official Transformers model at a pinned checkpoint and library version, so the reference must be reproducible. tools/zig/glm_ocr_reference.py has four commands: capture runs the model on one HIP GPU and keeps the tensors, tokens and runtime versions that native comparisons read; processor writes the processor and RoPE fixtures without a GPU; attention captures SDPA outputs for the AOTriton checks; and examples prepares the model-card tasks and linked images. Each command imports its heavy packages only when it runs.

The image codecs apply one shared source size limit, but GLM-OCR must decode a full linked page image unchanged before it resizes it. A new picture.Limits budget and decodeWithLimits entry points let a caller set its own side and pixel limits, which the PNG, JPEG and WebP decoders check before they allocate pixels; decode keeps the shared defaults and the decompression bomb bound still applies. The GLM-OCR processor passes its own budget, and a new test-picture-host step covers the caller limits.

GLM-OCR's CogViT vision tower and learned connector now run on HIP with owned scratch memory: MIOpen runs the patch convolution, rocBLAS the projections, a standalone AOTriton forward kernel the attention, and new pointwise kernels the norms, rotary embeddings, activations and layout steps. The convolution always takes MIOpen's GEMM solution from the returned candidates, so GPU timing cannot change its reduction order between requests. The new kernels join the HIP contract registry with gates and launch records, device tests compare patch, attention and encoder stages with the reference, and the AOTriton license joins the third-party notices.

The text decoder now runs on HIP from native image and token rows. A new hip.blas_fixed wrapper keeps one hipBLASLt single-row algorithm for each projection shape, so a row reduces in the same order at every request width, and KV restore, shared decoding and verification windows keep the reference arithmetic. Device tests compare decoder states and tokens with the captured references, and the shared BLAS device test now also checks the fixed path against exact integer dots.

Native open settings lived in the Qwen3.5 HIP engine, and only a family's name could tell the lane host that it accepts images. The settings move into the shared hip_serve layer with an optional vision field, and the native registry passes a loaded family's image offer to the lane host, so another native HIP family can open with the same settings and accept images without a special case. test-hip-native-options-host and test-hip-family-offers-host cover both paths without a device.

Draft verification needs logits for several rows at once. The decoder now keeps up to 32 rows of verification logits on the GPU for the shared drawer and computes every row with the same single-row vocabulary product, so batched verification returns the bits that one-row decoding returns. The operators can also load on a chosen device, and the device checks compare batched logits with single-row ones and prompt cuts, including cuts inside image rows, with a complete prefill.

GLM-OCR now registers as a native HIP family and serves chat requests through the shared lane core, GPU drawer and copy drafts, with bounded image preparation and admission against live device memory. The model never writes a reasoning block although its chat template has an enable_thinking switch, so the family declares reasoning false and the server keeps every reply in content while the prompt still follows enable_thinking. The family refuses MTP drafts and tensor ranks before it loads, and test-glm-ocr-native-host covers those refusals.

OCR clients often send the same page image again, and a fresh prefill of its image rows repeats the costliest prompt work. GLM-OCR prompt snapshots now keep the full prompt arithmetic shape and one keyed prepared image row, so a resumed prompt computes the same bytes as a complete prefill and keeps its image identity. The family opts in to the shared repeat_keyed cache rule, so a repeated image request earns a repeat cut, and a device check compares resumed decoding with a complete prefill. docs/recipes/glm-ocr-hip.md shows how to build, serve and check the model.

The registry test from ashhart#1933 checks the hook on two anonymous namespaces, so it does not name GLM-OCR. The GLM-OCR flag test now asserts that the registered family declares no reasoning block, which keeps its OCR text in message content. Only this family's assertion is added, so the other OCR families do not conflict with it.

The oracle hashed config.json and the model class file but accepted any checkpoint or Transformers tree. It now holds the SHA-256 of every non-weight checkpoint source at 2e85a6284 and of the GLM-OCR model, configuration and processor files in Transformers 5.5.4, and refuses a difference before a capture or fixture run. A new sources command checks both trees and prints the manifest without loading the model.

ashhart · 2026-11-03T12:12:00Z

Thank you @laudney. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.

ashhart · 2026-11-03T18:55:28Z

Thanks @laudney! This is in TensorFold 1.2.2, out now: https://github.com/ashhart/TensorFold/releases/tag/v1.2.2

laudney mentioned this pull request Nov 2, 2026

Native OCR models on HIP: dots.mocr, GLM-OCR, LightOnOCR-3 and Nemotron Parse, CUDA to follow (tracker) #1927

Open

6 of 11 tasks

AbrahamCete mentioned this pull request Nov 2, 2026

perf(hip): the decode softmax stats and partial sums load sixteen ahead #1961

Merged

9 tasks done

ashhart added a commit to laudney/TensorFold that referenced this pull request Nov 2, 2026

changelog: ashhart#1940 under Unreleased

0a47e4d

ashhart force-pushed the pr-ocr-glm-ocr branch from 1aab5dd to 0a47e4d Compare November 2, 2026 21:38

ashhart added a commit to laudney/TensorFold that referenced this pull request Nov 2, 2026

changelog: ashhart#1940 under Unreleased

6f83312

ashhart force-pushed the pr-ocr-glm-ocr branch from 0a47e4d to 6f83312 Compare November 2, 2026 21:58

laudney mentioned this pull request Nov 2, 2026

LightOnOCR-3 (0.8B, 1B, 4B) on native HIP, with the model card's oracle #1965

Open

ashhart mentioned this pull request Nov 2, 2026

dots.mocr on HIP: the publisher's BF16 OCR checkpoint, token-exact against its model card #1938

Merged

ashhart added a commit to laudney/TensorFold that referenced this pull request Nov 2, 2026

changelog: ashhart#1940 under Unreleased

7ac0ae5

ashhart force-pushed the pr-ocr-glm-ocr branch from 6f83312 to 7ac0ae5 Compare November 2, 2026 22:33

laudney mentioned this pull request Nov 2, 2026

Nemotron Parse on HIP: NVIDIA's checkpoint, token-exact against its HF example #1941

Open

ashhart force-pushed the pr-ocr-glm-ocr branch from 7ac0ae5 to 133348b Compare November 3, 2026 04:27

ashhart added a commit to laudney/TensorFold that referenced this pull request Nov 3, 2026

changelog: ashhart#1940 under Unreleased

133348b

ashhart added a commit to laudney/TensorFold that referenced this pull request Nov 3, 2026

changelog: ashhart#1940 under Unreleased

6010aca

ashhart force-pushed the pr-ocr-glm-ocr branch 2 times, most recently from 6010aca to 474ef4a Compare November 3, 2026 11:43

ashhart added a commit to laudney/TensorFold that referenced this pull request Nov 3, 2026

changelog: ashhart#1940 under Unreleased

474ef4a

laudney added 11 commits November 3, 2026 12:11

laudney and others added 2 commits November 3, 2026 12:11

changelog: ashhart#1940 under Unreleased

216a68e

ashhart force-pushed the pr-ocr-glm-ocr branch from 474ef4a to 216a68e Compare November 3, 2026 12:11

ashhart merged commit c7952f3 into ashhart:main Nov 3, 2026
0 of 7 checks passed

laudney mentioned this pull request Nov 3, 2026

feat(hip): one host protocol for the HIP OCR families #2040

Merged

Navigation Menu

Repository navigation

GLM-OCR on HIP: the publisher's BF16 checkpoint, exact against its six model-card tasks - #1940

GLM-OCR on HIP: the publisher's BF16 checkpoint, exact against its six model-card tasks#1940
ashhart merged 13 commits into
ashhart:mainfrom
laudney:pr-ocr-glm-ocr

laudney commented Nov 2, 2026 •

edited

Loading

Uh oh!

ashhart commented Nov 2, 2026

Uh oh!

laudney commented Nov 2, 2026

Uh oh!

Uh oh!

ashhart commented Nov 3, 2026

Uh oh!

ashhart commented Nov 3, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

laudney commented Nov 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Commits

Runtime libraries (optional, loaded at run time)

Known limits

Gates

Uh oh!

ashhart commented Nov 2, 2026

Uh oh!

laudney commented Nov 2, 2026

Device receipt

Kernel contract and limits (the two new commits change no kernel)

Uh oh!

Uh oh!

ashhart commented Nov 3, 2026

Uh oh!

ashhart commented Nov 3, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

laudney commented Nov 2, 2026 •

edited

Loading