Add files using upload-large-folder tool

.gitattributes

@@ -33,3 +33,10 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+gpu-pipelined/clef_flash_decode_fp16_pf64/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+gpu-pipelined/clef_flash_decode_int8mix_pf64/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+gpu-pipelined/clef_flash_head_bucket_fp16w32/clef_flash_head_bucket_fp16w32.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
+gpu-pipelined/clef_flash_g448_vision_fp16w32/clef_flash_g448_vision_fp16w32.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
+gpu-pipelined/clef_flash_g256_vision_fp16w32/clef_flash_g256_vision_fp16w32.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
+gpu-pipelined/clef_flash_decode_int8mix_pf64/clef_flash_decode_int8mix_pf64.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
+gpu-pipelined/clef_flash_decode_fp16_pf64/clef_flash_decode_fp16_pf64.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Alibaba Cloud
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -0,0 +1,474 @@
+---
+license: apache-2.0
+base_model: Cloudflare/clef-flash
+base_model_relation: quantized
+pipeline_tag: image-text-to-text
+library_name: coreai
+language:
+- en
+tags:
+- coreai
+- core-ai
+- apple
+- on-device
+- decision-model
+- systemone
+- typed-decisions
+- vision-language
+- structured-output
+---
+
+# clef-flash — Core AI
+
+Apple Core AI (`.aimodel`) conversion of [Cloudflare/clef-flash](https://huggingface.co/Cloudflare/clef-flash/tree/17f0b0ad64efb65d273590632833508766b2aae6) (revision `17f0b0a`, Apache-2.0): a state (text, JSON or an image) and typed questions in, the probability of every option of every question out, from one prefill.
+
+| Variant | Path | Size | Requires | Tested on |
+|---|---|---:|---|---|
+| Decoder, fp16 (default) | `gpu-pipelined/clef_flash_decode_fp16_pf64/` | 15,900 MB | macOS 27 | M4 Max, macOS 27.0 (26A428), 2026-10-03 |
+| Decoder, int8 except layers 0–11 | `gpu-pipelined/clef_flash_decode_int8mix_pf64/` | 11,847 MB | macOS 27 | M4 Max, macOS 27.0 (26A428), 2026-10-03 |
+| Joint schema head (graph) | `gpu-pipelined/clef_flash_head_bucket_fp16w32/` | 244 MB | macOS 27 | M4 Max, macOS 27.0 (26A428), 2026-10-03 |
+| lm_head gather table (host) | `host/lm_head_fp16.bin` | 2,034 MB | — | M4 Max, macOS 27.0 (26A428), 2026-10-03 |
+| Vision tower, 256×256 → 64 image rows | `gpu-pipelined/clef_flash_g256_vision_fp16w32/` | 909 MB | macOS 27 | M4 Max, macOS 27.0 (26A428), 2026-10-03 |
+| Vision tower, 448×448 → 196 image rows | `gpu-pipelined/clef_flash_g448_vision_fp16w32/` | 912 MB | macOS 27 | M4 Max, macOS 27.0 (26A428), 2026-10-03 |
+
+A decision needs one decoder, the head and the table; an image adds one tower. Mac only: the decoder alone is 15.9 GB (fp16) or 11.8 GB (int8mix). On the M4 Max a 436-token decision takes 0.95 s in Swift ([measured below](#time-per-decision)).
+
+```swift
+import CoreAI
+import ClefFlash
+
+let root = URL(filePath: "clef-flash-CoreAI")                        // a download of the HF repo
+let gp = root.appending(path: "gpu-pipelined")
+var decoderOptions = SpecializationOptions(preferredComputeUnitKind: .gpu)
+decoderOptions.expectFrequentReshapes = true
+let gpu = SpecializationOptions(preferredComputeUnitKind: .gpu)
+let decider = try await ClefDecider(
+    assets: .init(decoderBundle: gp.appending(path: "clef_flash_decode_fp16_pf64"),
+                  head: gp.appending(path: "clef_flash_head_bucket_fp16w32"),
+                  table: root.appending(path: "host/lm_head_fp16.bin"),
+                  towers: [.g448: gp.appending(path: "clef_flash_g448_vision_fp16w32/clef_flash_g448_vision_fp16w32.aimodel")]),
+    decoderOptions: decoderOptions, headOptions: gpu, towerOptions: gpu)   // nil decoder / headAsset = the .aimodel (JIT)
+let request = try SystemOneRequest(data: try Data(contentsOf: requestURL))
+let response = try await decider.decide(request: request, image: cgImage, grid: .g448)   // image: nil = text only
+print(PythonJSON.dumps(response, sortKeys: false))
+// {"model":"clef-flash","answers":{"department":{"type":"choice","choice":…,"confidence":…,"probabilities":{…}},…},
+//  "usage":{"input_tokens":…,"output_tokens":0}}
+```
+
+The snippet uses the [`ClefFlash`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/ClefFlash) Swift package from the model zoo. The zoo card and the gate scripts live in [coreai-model-zoo](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/README.md); the rest of this page is the same text with repository links.
+
+Source [Cloudflare/clef-flash](https://huggingface.co/Cloudflare/clef-flash/tree/17f0b0ad64efb65d273590632833508766b2aae6) (revision `17f0b0a`) · base Qwen/Qwen3.5-9B
+
+A **decision model**. Give it a state (text, JSON or an image) and typed questions: `noul` (true /
+false), `choice` (named options) or `score` (ordered levels). It returns a probability for every
+allowed option of every question from one prefill. It never generates text. Requests and responses
+use the SystemOne-compatible request shape: `{model, state, questions}` in, `{model, answers, usage}`
+out, one typed decision per question.
+
+Cloudflare post-trained it from Qwen3.5-9B. Their announcement: "By freezing Qwen3.8-27B for Clef and
+Qwen3.5-9B for Clef-flash, we jointly optimized the routing head alongside rank-256 low-rank
+adapters." The checkpoint ships the backbone (24 Gated DeltaNet + 8 full-attention layers, a
+27-block vision tower) as sharded safetensors, plus a 121.8M-parameter joint schema head that reads
+the backbone's final hidden states. The author's card reports benchmark results; none of them are
+re-measured here.
+
+This port runs that readout as three Core AI graphs and one host table. The **decoder** is the
+Qwen3.5 hybrid with token ids in, the image rows as a static input and M-RoPE derived inside the
+graph. It returns the final-norm hidden state at every position and has no vocabulary head. One
+call takes 64 tokens. Two bundles: fp16 (15.9 GB, the default) and int8mix (int8 per block of 32
+except layers 0–11, 11.8 GB). The **vision tower** is baked at a fixed grid, `g256` (64 image
+tokens) or `g448` (196), fp16 weights read through a cast with fp32 math (909 / 912 MB). The
+**head** is the author's joint schema head written as plain matrices, 244 MB, with four static key
+lengths. The untied `lm_head.weight` stays on the host as a 2.03 GB fp16 gather table for the head's
+lexical term. The gate is probability parity with the author's own fp32 code, on every option of
+every question, on a self-made fixture and on a held-out record set written after the int8 layers
+were chosen.
+
+The port runs on the Mac GPU. iPhone is not a target: the decoder alone is 15.9 GB (fp16) or
+11.8 GB (int8mix), and an iOS app had about 6.4 GB available on an iPhone 18 Pro with the
+increased-memory-limit entitlement ([decider-2b-vision](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/decider-2b-vision/README.md)).
+
+## Readout contract
+
+One record per request, every question answered in the same pass. The ids are the checkpoint's
+`encode_record()`: prefix + [image block] + state + schema + suffix, each piece tokenized on its own
+without special tokens, then concatenated:
+
+```
+<|im_start|>system\nRead the complete state and schema. Decide every field jointly. Each answer must be exactly one of that field's allowed options.<|im_end|>\n<|im_start|>user\nSTATE:\n[<|vision_start|><N image tokens><|vision_end|>\n]<state>\n\nSCHEMA FIELDS:\n\nFIELD 1\nID: <question id>\nTYPE: <type>\nINSTRUCTION: <instructions>\nALLOWED OPTIONS:\nOPTION 1: {"description":"…","option_id":"…"}\nOPTION 2: …\nEND FIELD\n\nFIELD 2\n…END FIELD\n\n<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\nJOINT SCHEMA DECISIONS:
+```
+
+- The prefix is 36 tokens for every record and the suffix 18, ending in `:` (25). The image block
+  starts at index 36: `<|vision_start|>` (248053), N image tokens, `<|vision_end|>` (248054), `\n`.
+- `render()`: a string as is, anything else `json.dumps(sort_keys=True, ensure_ascii=False,
+  separators=(",", ":"))`. The state is rendered this way; so is each question's instructions (the
+  question id when there are none) and each option's `{"option_id", "description"}`.
+- A question's span is its rendered instructions; an option's span is its rendered object. The head
+  averages the hidden rows over each span, reads the last token as a global vector, and averages the
+  `lm_head.weight` rows of each option span's ids (the lexical term).
+- Option order: `noul` = (true, false), with the default descriptions "The proposition is true or the
+  answer is yes." / "…false or the answer is no." unless the request overrides them; `choice` = option
+  ids sorted as strings; `score` = levels 0..n−1.
+- Probabilities: a float32 softmax per question over that question's options. Response: `noul` →
+  p(true); `choice` → the argmax id, its p, every p; `score` → Σ level·p, the largest p, the legend,
+  every p; all rounded to 4 decimals. `usage` = `{input_tokens: T, output_tokens: 0}`.
+- Image: one image per request. RGB → Pillow BICUBIC resize to 256×256 or 448×448 (the aspect ratio is
+  not kept) → /255 → (x − 0.5) / 0.5 → merge-block-major patches `[4G², 1536]`, the frame repeated at
+  both temporal slots. The author's processor picks a grid per image (at least 256×256);
+  [what a fixed grid costs](#what-a-fixed-grid-costs-the-authors-code-only) is measured below. The
+  graph takes the image tokens as ids V + k (V = 248,320, k row-major over the merged grid), so each
+  token names its tower row.
+- Text-only record: no image block, the same bundle, zero image rows and `rope_shift_start` = 2³⁰.
+- Length: a record and its padded last call must fit 4,096 tokens, the image block included. The
+  author's own cut at 16,384 tokens is never reached.
+
+`conversion/clef_flash/oracle_clef.py` runs the checkpoint's own `joint_schema_model.py`
+(`load_release_model()` + `systemone()`, fp32, CPU) and records the reference:
+[`fixtures-clef-flash.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/fixtures-clef-flash.json). It holds 216 records with their requests,
+ids, spans, option logits and probabilities: 186 written for the port (16 text states, three of
+them about 2,000 tokens; 12 JSON states; 10 images drawn by `make_fixtures.py`; 4 CC0 photographs;
+the 144 `choice` items of SemIf authored144) and 30 written later as a held-out set. Each image
+record runs at `g256`, `g448` and `native` (the processor's own grid): 214 runs, 405 questions. The
+gates use 213 runs and 401 questions: `photo_01` at `native` has 1,200 image rows, more than the
+graph's 1,024-row buffer. Four questions have an oracle top-2 margin of 0.02 or less (near-ties) and
+are listed apart. The held-out set runs text, `g256` and `g448`: 40 runs, 188 questions, 2 near-ties.
+
+## Core AI shape
+
+**Decoder.** `conversion/clef_flash/qwen3_5_clef_decoder.py`: decider-2b-vision's ids-input Qwen3.5
+VL decoder (`conversion/decider_vision/qwen3_5_vl_pipelined.py`) at 9B, without the vocabulary head,
+returning the hidden state at every position. One function, `main`, at a static 64 tokens; the Gated
+DeltaNet recurrence is unrolled in the graph.
+
+| | name | shape, type |
+|---|---|---|
+| inputs | `input_ids` | [1, 64] int32 |
+| | `position_ids` | [1, seq] int32, the ramp 0 .. seq − 1 |
+| static inputs | `image_embeds` | [1024, 4096] fp16: tower rows 0..N−1, the rest zero |
+| | `image_rc` | [1024, 2] int32: (k // W, k % W) for image token k |
+| | `rope_shift_start` | [1] int32: 37 + N, the `<\|vision_end\|>` index |
+| | `rope_shift_amount` | [1] int32: N − max(H, W) |
+| states | `keyCache`, `valueCache` | [8, 1, 4, ctx, 256] fp16, ctx up to 4,096 |
+| | `convState`, `recState` | [24, 1, 8192, 3], [24, 1, 32, 128, 128] fp16 |
+| output | `hidden` | [1, 64, 4096] fp16, every position |
+
+Per record, from zeroed states: ⌈T / 64⌉ calls, call k with ids[64k : 64k + 64] and position_ids
+0..64k + 63; the last call is padded with `<|endoftext|>` (248044) and its padded rows are dropped.
+The rows of every call, cut to T, are the backbone's final-norm `last_hidden_state [T, 4096]`. The
+bundle's `metadata.json` says `kind: decision-backbone` and carries this order, the prompt pieces,
+the span rules, the option order and the response shape.
+
+**Towers.** The overlay's `qwen3_5_vision.Qwen3_5VisionEncoder` through
+`conversion/export_qwen38vl_pipelined.py --skip-decoder --vision-dtype fp16w32`, baked at 8×8 or
+14×14 merged tokens: patches `[4G², 1536]` f32 → `image_embeds [G², 4096]` f32.
+
+**Head.** `conversion/clef_flash/clef_head.py`, `ClefHeadGraph`: the author's `JointSchemaHead` (its
+own submodules, `load_state_dict(strict=True)`) for one record, its Python loops over questions and
+options written as matrices, fp16 weights read through a cast with fp32 math. Ten inputs: `hidden
+[T, 4096]` f32, `key_valid [T]`, `q_avg [Q, T]` and `o_avg [O, T]` (1/len over each span), `g_avg
+[1, T]` (the last real token), `lexical [O, 4096]`, `member [Q, O]`, `type_ids [Q]` int32, `q_valid
+[Q]`, `o_valid [O]`; output `logits [O]` f32. Four functions share the weights, `t512`, `t1024`,
+`t2048`, `t4096`; the host pads T to the smallest that holds it, Q to 16 and O to 128, with zero rows
+and valid flags 0. Its `metadata.json` says `kind: decision-head`. Above 2,048 keys the head's
+attention runs in key blocks of at most 2,048 (see [Precision](#precision)).
+
+**Host.** Builds the ids and spans, preprocesses the image, runs the decoder calls, converts the
+hidden rows to f32 and builds the ten head arrays (the lexical rows from `host/lm_head_fp16.bin`,
+`[248320, 4096]` fp16, row t at byte offset t × 8192), then the per-question softmax and the
+response. `conversion/clef_flash/decide.py` is the Python reference; `apps/ClefFlash` is the Swift
+one.
+
+No engine is involved: the hosts drive the low-level runtime (`AIModel` + `loadFunction`, four
+zeroed states per record). No runtime patch.
+
+## Measured (Apple M4 Max, macOS 27.0 26A428, 2026-10-03)
+
+The bar, fixed before any graph ran: the argmax equal to the oracle's on every question whose oracle
+top-2 margin is above 0.02 (near-ties listed apart), max |Δp| ≤ 0.02 over every option of every
+question, the mean over runs of each run's mean |Δp| ≤ 0.002, and every process re-running its first
+run bit for bit. The Python gates load AOT `.aimodelc` assets (`coreai-build compile … --platform
+macOS --preferred-compute gpu --architecture h16c`, `--expect-frequent-reshapes` for the decoder) with
+`SpecializationOptions.default()`; their times are on a GPU shared with other work.
+
+### Before any graph: fp32 torch
+
+The decoder module in fp32 on the CPU, driven like the graph (chunks from zeroed states, the unrolled
+recurrence, M-RoPE in the module, the oracle's own tower rows), read through the author's head: 214
+runs, argmax 405/405, max |Δp| 9.6e-6, lowest per-position hidden cosine 0.99999939; its M-RoPE
+planes equal the oracle's on 214/214 runs. With zero image inputs its hidden state equals the
+overlay's plain Qwen3.5 text decoder bit for bit (3/3 runs). Zeroing the image rows changes 2 of 3
+answers on `img_07` at `g448` and on `img_01` at `g256` (max |Δp| 0.66 / 0.75)
+([`gate-clef-flash-torch-parity.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/gate-clef-flash-torch-parity.json)).
+
+The host spec (`conversion/clef_flash/host.py`) rebuilds every oracle run's ids, image offset, spans,
+option order and M-RoPE planes: 242/242 runs with two tokenizer implementations. Its Pillow-resized
+patches equal the processor's `pixel_values` bit for bit on all 56 fixed-grid runs.
+
+### Vision tower
+
+The shipped fp16w32 tower, AOT for the Mac GPU, fed the host's patches, against the oracle's fp32
+tower rows: `g256` 14/14 images, worst image cosine 0.9999999994, worst row 0.99999995; `g448`
+14/14, 0.9999999994 and 0.99999995; 31 / 79 ms per image on the shared GPU. Patches in raster order
+instead of merge-block order miss on every image ([`gate-clef-flash-tower.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/gate-clef-flash-tower.json)).
+
+### Decoder alone on the Mac GPU: 213 fixture runs
+
+The oracle's ids and fp32 image rows in, the fp16 hidden read through the author's fp32 head:
+
+| decoder | chunk | argmax (margin > 0.02) | near-ties agreeing | max \|Δp\| | mean of run means | bar |
+|---|---:|---:|---:|---:|---:|---|
+| **fp16 (default)** | 64 | **397/397** | **4/4** | **0.0112** | **0.00054** | **PASS** |
+| **int8mix: int8 except layers 0–11** | 64 | **397/397** | **4/4** | **0.0153** | **0.00089** | **PASS** |
+| int8lin: every decoder linear int8 | 64 | 397/397 | 3/4 | 0.0748 | 0.00192 | FAIL |
+| fp16 | 16 | 397/397 | 4/4 | 0.0127 | 0.00053 | PASS |
+
+int8 is per block of 32 (`symmetric_with_clipping`, weights only); the embedding table, the Gated
+DeltaNet conv1d and every norm stay fp16. int8lin misses on 9 of 213 runs, worst a SemIf record at
+0.0748 (fp16 on the same record: 0.0105), and flips the `img_06` `g448` near-tie (margin 0.0175).
+The red arm, the image rows zeroed on `img_07` `g448` and `img_01` `g256`, moves p by 0.66 and 0.75
+and changes one or two of the three answers on every bundle that ran it. Transcripts:
+`gate-clef-flash-readout-<bundle>.json`.
+
+### The head graph
+
+The head graph alone on the oracle's fp32 hidden: 214 runs, argmax 405/405, max |Δp| 1.8e-7; held
+out, 40 runs, 188/188, 1.8e-7. End to end in Python from the fp16 decoder's hidden through the host
+arrays and the head graph: the same values as the author's fp32 head on that hidden (0.0112 fixture,
+0.0085 held out), within 1.8e-7 / 2.4e-7. Three perturbed arms on 4 runs each move p: span rows one
+token late by 0.003–0.056, the lexical term zeroed by 0.017–0.055, option membership rotated by
+0.051–0.348. `decide.py` from the raw request, on 12 fixture and 12 held-out requests: ids and spans
+equal the oracle's 24/24, every choice equal (18/18, 32/32), the 4-decimal values within 0.0082
+([`gate-clef-flash-head.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/gate-clef-flash-head.json)).
+
+### Held out: 40 runs that no choice was made on
+
+30 records written after the int8 layers were chosen, with new names, drawings and photographs and a
+novelty check against the fixture (no shared record, image, name or 8-word run), and their own fp32
+oracle run:
+
+| decoder | argmax (margin > 0.02) | near-ties agreeing | max \|Δp\| | mean of run means | bar |
+|---|---:|---:|---:|---:|---|
+| fp16 (default) | 186/186 | 2/2 | 0.0085 | 0.00036 | PASS |
+| int8mix | 186/186 | 1/2 | 0.0171 | 0.00063 | PASS |
+
+int8mix flips the `ho_t07` near-tie (oracle margin 0.0117: p 0.4102 / 0.3985, int8mix 0.3979 /
+0.4059) ([`gate-clef-flash-heldout.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/gate-clef-flash-heldout.json)).
+
+### What a fixed grid costs (the author's code only)
+
+`conversion/clef_flash/grid_price.py` runs the checkpoint's own code on the 14 image records (50
+questions), the image resized to a square grid first, against the processor's own grid (`native`),
+fp32 on the CPU:
+
+| arm | image tokens | argmax = native | max \|Δp\| vs native | mean \|Δp\| vs native | gold correct |
+|---|---:|---:|---:|---:|---:|
+| native | 412 (mean) | — | — | — | 35/38 |
+| `g256` | 64 | 46/50 | 0.959 | 0.039 | 33/38 |
+| `g448` | 196 | 48/50 | 0.436 | 0.018 | 36/38 |
+| 672×672 | 441 | 48/50 | 0.350 | 0.016 | 36/38 |
+| 896×896 | 784 | 48/50 | 0.244 | 0.013 | 36/38 |
+
+The two larger grids agree with `native` no more often than `g448`, so only `g256` and `g448` ship
+([`gate-clef-flash-variants.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/gate-clef-flash-variants.json)).
+
+### Swift ([`apps/ClefFlash`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/ClefFlash), Mac, Release)
+
+Images through the Swift tower; the 13 `native` fixture runs fed the oracle's rows:
+
+| decoder asset | fixture 213: argmax / near-ties / max \|Δp\| / mean | held-out 40 |
+|---|---|---|
+| fp16, AOT `.aimodelc` | 397/397, 4/4, 0.0122, 0.00054 | 186/186, 2/2, 0.0082, 0.00037 |
+| fp16, JIT: the `.aimodel` specialized by the Swift runtime | the same values | the same values |
+| int8mix, AOT | 397/397, 4/4, 0.0153, 0.00089 | 186/186, 1/2, 0.0171, 0.00065 |
+| int8mix, JIT | 397/397, 4/4, 0.0148, 0.00089 | 186/186, 1/2, 0.0173, 0.00065 |
+
+- On identical inputs Swift computes what Python computes: ids and spans from the raw request
+  253/253 runs; the renderer = CPython's `json.dumps` / `repr` / `round(x, 4)` on 2,204 values and
+  199,994 doubles; tiles and patches = Pillow and the oracle's `pixel_values` on 48/48; tower rows =
+  the Python runtime's on 48/48; with fp16 AOT, hidden rows and head logits bit-equal on 185 + 20
+  runs and on the 48 runs fed the oracle's image rows (int8mix AOT: hidden rows bit-equal on the
+  same runs), and responses = `decide.py`'s on 24/24. The image runs differ from the Python table
+  above only through the tower rows (the Python readout fed the oracle's).
+- JIT against AOT: fp16 bit-equal on 253/253 runs (hidden, logits, responses). int8mix is not
+  bit-equal on any run: max |Δp| 0.0014, max |Δlogit| 0.020, argmax 589/589.
+- The JIT's first specialization: fp16 56.1 s (the decoder 50.0 s), the runtime's cache +31.8 GB,
+  4.5 s from the cache; int8mix 42.3 s (39.2 s), +17.0 GB, 3.0 s.
+- Changing one word of one question moves the ids and spans off the oracle's and moves p (up to
+  0.0056 on `own_t01`).
+
+Transcript: [`gate-clef-flash-swift.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/gate-clef-flash-swift.json).
+
+### Time per decision
+
+Swift, Release CLI, AOT assets, the machine-wide GPU lock taken with the GPU at 0 %. Two passes per
+bundle, each in its own process, one warm-up decision, then 11 decisions × 3; wall time from the file
+read to the response, ms:
+
+| record | T | decoder calls | head | fp16 median (min–max) | int8mix median | tokenize | image decode + resize + patches | tower | decoder | head |
+|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|
+| SemIf `a3f18f3a` | 185 | 3 | t512 | 411 (411–416) | 412 | 0.8 | – | – | 401 | 7.4 |
+| SemIf `6149a17b` | 200 | 4 | t512 | 545 (544–547) | 547 | 0.9 | – | – | 535 | 7.4 |
+| `own_t01` | 436 | 7 | t512 | 949 (947–949) | 951 | 1.9 | – | – | 938 | 7.4 |
+| `own_t11` | 497 | 8 | t512 | 1,084 (1,082–1,085) | 1,085 | 2.2 | – | – | 1,073 | 7.4 |
+| `own_t12` | 967 | 16 | t1024 | 2,165 (2,157–2,167) | 2,168 | 3.9 | – | – | 2,147 | 12.0 |
+| `own_t16` | 2,494 | 39 | t4096 | 5,289 (5,280–5,293) | 5,291 | 10.8 | – | – | 5,249 | 27.7 |
+| `own_t15` | 2,603 | 41 | t4096 | 5,563 (5,551–5,576) | 5,572 | 11.7 | – | – | 5,522 | 27.5 |
+| `photo_02` `g256` | 499 | 8 | t512 | 1,138 (1,136–1,142) | 1,139 | 1.7 | 11.6 | 39.6 | 1,073 | 10.0 |
+| `img_04` `g256` | 558 | 9 | t1024 | 1,273 (1,271–1,276) | 1,276 | 1.7 | 8.9 | 40.0 | 1,207 | 12.3 |
+| `photo_02` `g448` | 631 | 10 | t1024 | 1,457 (1,455–1,459) | 1,459 | 1.8 | 13.4 | 87.1 | 1,342 | 12.1 |
+| `img_04` `g448` | 690 | 11 | t1024 | 1,586 (1,584–1,600) | 1,587 | 2.0 | 11.3 | 87.2 | 1,475 | 9.9 |
+
+The stage columns are the fp16 medians. One decoder call (64 tokens) takes 134.2 ms with fp16 (936
+calls, 132.6–145.3) and 134.3 ms with int8mix (132.7–149.8). Load from the runtime's cache 3.3 s
+(fp16); int8mix 13.3 s in its first pass, right after the fp16 passes, and 2.9 s in the second. The
+first decision of a process takes 3.4–3.7 s, its first decoder call 2.6–2.8 s. Zeroing the four
+states takes 1.3 ms. Transcript: [`gate-clef-flash-timing.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/gate-clef-flash-timing.json).
+
+## Precision
+
+**int8.** int8 over every decoder linear misses the bar (int8lin above). An fp32 torch bisect with
+the exporter's own int8 weights (`conversion/clef_flash/int8_bisect_torch.py`, the six runs int8lin
+misses worst) puts the error in layers 0–15: those 16 layers in fp16 leave the worst run at 0.0023,
+layers 16–31 in fp16 change nothing (0.0845, all int8: 0.0842). The error is spread out and partly
+cancels: one kind of linear alone in int8 is already large (Gated DeltaNet projections 0.0752, MLP
+0.0342, attention 0.0230), no single fp16 layer brings the worst run below 0.0492, and some single
+layers make it worse (layer 3: 0.1164).
+
+The rule was written before any bisect result: worst run ≤ 0.010, no run worse than int8lin, at most
+six fp16 layers, the smallest set. No set met it; the ranked sets of two to six layers stayed at
+0.0260–0.0350. Contiguous fp16 blocks, run after the rule's search and outside it:
+
+| fp16 layers | 0–7 | 0–8 | 0–9 | 0–10 | 0–11 | 4–15 | 8–15 | 12–15 |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| worst run max \|Δp\| | 0.0568 | 0.0509 | 0.0232 | 0.0116 | 0.0069 | 0.0207 | 0.0296 | 0.0939 |
+
+int8mix keeps layers 0–11 fp16, twelve layers, outside the rule's cap. It was chosen on the
+fixture; the held-out records above are its test (186/186, max 0.0171). It saves 4.05 GB against
+fp16 (`main.mlirb` 11,826,520,241 vs 15,880,253,283 bytes) and no time: on the same 202 runs a call
+takes 134.4 ms (int8mix), 134.3 (int8lin) and 134.2 (fp16); why int8 buys no time is not isolated
+([`gate-clef-flash-int8-bisect.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/gate-clef-flash-int8-bisect.json)).
+
+**fp16.** The fp16 decoder passes with room, but a few hidden rows drift: 50 of 73,343 positions
+have a cosine below 0.99 against the oracle (34 of them image rows), 6 below 0.9, one below 0.5,
+0.41 at an image row of `photo_04` `g448`, a run whose max |Δp| is 1.1e-4. The fp16 module in torch
+on MPS finds the same row hardest but keeps it at 0.877. Computing the full-attention linears and SDPA
+in fp32 does not lift it (38 runs: lowest 0.378 vs 0.407, 53 vs 48 positions below 0.99, max |Δp|
+0.0075 in both, 7% slower per call). Which op opens the gap is not isolated
+([`gate-clef-flash-variants.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/gate-clef-flash-variants.json)).
+
+**Chunk width.** 64 tokens per call: 134.2 ms per call, 2.26 ms per token, against 58.8 ms and 3.77
+ms per token at 16 (202 runs, shared GPU), the argmax equal on all 213 runs and p within 0.0043. 128
+tokens per call bought nothing on 19 runs: 2.14 vs 2.19 ms per token, 1.227 vs 1.210 s per run with
+the longer padding.
+
+**Head attention over long inputs.** On this GPU stack the plain `softmax(q kᵀ · s + b) v` chain
+compiled for the head gave wrong values, different on every call, from 4,032 keys on (exact at
+3,584), while each op alone was exact at 4,096. Keys in blocks of at most 2,048 with one shared max
+are exact at 4,032 and 4,096 (max |d| 5.7e-7), so the `t4096` function uses that form. The decoder's
+own attention showed nothing of the kind on a 4,050-token row whose last calls attend 4,032 and 4,096
+keys: one position of 4,050 is below a cosine of 0.99 (index 1,300, 0.9585), max |Δp| 6.1e-4
+([`gate-clef-flash-head.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/gate-clef-flash-head.json)).
+
+**The lm_head table.** fp16 rows move p by at most 1.8e-7 (the fp32 table's own re-run floor); int8 per
+row would move it by 4.4e-4 and save 1.0 GB. The table ships in fp16.
+
+## ⬇️ Bundle
+
+Five folders under `gpu-pipelined/` and the table under `host/`. A decision needs one decoder, the head and the
+table; an image adds one tower.
+
+| folder | what | main.mlirb bytes | main.mlirb sha256 |
+|---|---|---:|---|
+| `gpu-pipelined/clef_flash_decode_fp16_pf64/` | decoder, fp16 (default): `.aimodel` + `metadata.json` + `tokenizer/` | 15,880,253,283 | `0b7dfa30…b9bea8` |
+| `gpu-pipelined/clef_flash_decode_int8mix_pf64/` | decoder, int8 except layers 0–11, same layout | 11,826,520,241 | `2c19a0e7…6a67b9` |
+| `gpu-pipelined/clef_flash_g256_vision_fp16w32/` | tower, 256×256 → 64 rows (`.aimodel`) | 909,207,876 | `b2adb039…64f31d` |
+| `gpu-pipelined/clef_flash_g448_vision_fp16w32/` | tower, 448×448 → 196 rows (`.aimodel`) | 911,945,013 | `04d8629b…3a1eaa` |
+| `gpu-pipelined/clef_flash_head_bucket_fp16w32/` | joint schema head, `t512`–`t4096` (`.aimodel` + `metadata.json`) | 243,978,965 | `51e512ad…6856fe` |
+
+| file | what | bytes | sha256 |
+|---|---|---:|---|
+| `host/lm_head_fp16.bin` | the untied `lm_head.weight`, `[248320, 4096]` fp16, row-major, no header | 2,034,237,440 | `bfccf00b…1ba36d` |
+| `host/lm_head_fp16.json` | its shape, dtype, source tensor and read-back check | 2,939 | `ef89e2fb…b75e86a` |
+
+`SHA256SUMS` in the repository lists every file. The Hub revision is filled in after the upload.
+
+No AOT asset ships: the Swift runtime specializes the `.aimodel` files correctly on the Mac (the JIT
+rows above); the decoder's AOT compile is 29.7 GB (fp16) / 25.7 GB (int8mix). The author's
+`joint_head.safetensors` is not included, because the head graph carries its weights. The source's
+`config.json` and `joint_head_config.json` are in the repository verbatim, with its Apache-2.0
+`LICENSE`.
+
+## Use it
+
+Swift: the snippet at the top of this page.
+
+The same from the CLI (`swift build -c release --package-path apps/ClefFlash` builds `clef-flash`):
+
+```bash
+R=clef-flash-CoreAI; G=$R/gpu-pipelined
+clef-flash ask --decoder-asset jit --decoder $G/clef_flash_decode_fp16_pf64 --head $G/clef_flash_head_bucket_fp16w32 \
+    --table $R/host/lm_head_fp16.bin \
+    --tower-g448 $G/clef_flash_g448_vision_fp16w32/clef_flash_g448_vision_fp16w32.aimodel \
+    --request req.json --image x.png --grid 448 --out resp.json
+```
+
+A request, in the SystemOne-compatible request shape:
+
+```json
+{"model": "clef-flash",
+ "state": "Our checkout started returning errors and orders are blocked.",
+ "questions": {
+   "department": {"type": "choice", "instructions": "Which team should handle the message?",
+                  "criteria": {"billing": "Payments or invoices", "technical": "Bugs or outages"}},
+   "urgency": {"type": "score", "criteria": ["Can wait", "This week", "Today"]},
+   "outage": {"type": "noul", "instructions": "Is a service down?"}}}
+```
+
+Python: `conversion/clef_flash/decide.py run --request req.json [--image x.png --grid 448] --out
+resp.json` is the gates' own read-out with `coreai.runtime`. It loads AOT `.aimodelc` assets only
+(compile each `.aimodel` with the `aot` line of [`recipe.toml`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/clef-flash/recipe.toml)); the Python gates never
+used the Python runtime's JIT.
+
+## Reproduce
+
+Environment: the zoo overlay venv (coreai-core 1.0.0b2, coreai-torch 0.4.1, torch 2.9.0,
+transformers 4.57.6), the source snapshot pinned at `17f0b0a` in `HF_HOME`, `HF_HUB_OFFLINE=1`,
+Xcode 27.0 RC. The author's code runs in its own venv (transformers 5.17.0, torch 2.9.0). The steps,
+in order, with every flag, are in [`conversion/clef_flash/README.md`](https://github.com/john-rocky/coreai-model-zoo/blob/main/conversion/clef_flash/README.md).
+
+```bash
+# towers: the Qwen3.8-27B vision exporter with the HF id and the grid set
+python conversion/export_qwen38vl_pipelined.py --hf-id Cloudflare/clef-flash --name clef_flash_g256 \
+    --grid-h 8 --grid-w 8 --skip-decoder --vision-dtype fp16w32 --out-dir <work>/_clefflash/exports
+python conversion/export_qwen38vl_pipelined.py --hf-id Cloudflare/clef-flash --name clef_flash_g448 \
+    --grid-h 14 --grid-w 14 --skip-decoder --vision-dtype fp16w32 --out-dir <work>/_clefflash/exports
+# decoders (--aot adds the h16c .aimodelc the Python gates load)
+python conversion/clef_flash/export_decoder.py fp16 --prefill-chunk 64 --aot
+python conversion/clef_flash/export_decoder.py int8mix --fp16-layers 0,1,2,3,4,5,6,7,8,9,10,11 --prefill-chunk 64 --aot
+# head and host table
+python conversion/clef_flash/export_head.py --shape bucket --weights fp16w32 --aot
+python conversion/clef_flash/export_lm_head_table.py
+```
+
+Port notes: [`knowledge/clef-flash-port.md`](https://github.com/john-rocky/coreai-model-zoo/blob/main/knowledge/clef-flash-port.md).
+
+## Other formats
+
+On the Hub (2026-10-03), not run here: MLX conversions with the joint head,
+`mlx-community/clef-flash-4bit` and `-8bit` and `TrevorJS/clef-flash-mlx-4bit` and `-8bit`; an
+OpenVINO conversion with the head, `meossistant/clef-flash-openvino`; GGUF conversions such as
+`bartowski/Cloudflare_clef-flash-GGUF`, which carry no joint head file. mlx-community's card on the
+backbone alone: "It is not a chat model — `mlx_vlm.generate`, `mlx_lm.generate`, and LM Studio will
+load the backbone but produce meaningless text." No other Core AI conversion of clef-flash was
+listed on the Hub on 2026-10-03.
+
+## License
+
+Source Apache-2.0 (Cloudflare/clef-flash, following its base Qwen/Qwen3.5-9B, also Apache-2.0); the
+bundles inherit it, and the Hugging Face repository carries the source's `LICENSE`. The author's
+`joint_schema_model.py` runs only in the oracle and the reference scripts at gate time; it is not
+part of the bundles. The fixture's drawn images are generated by `make_fixtures.py` (CC0-1.0); its
+photographs are Wikimedia Commons files marked CC0 (each file page, thumbnail URL and sha256 is in
+the fixture); its 144 SemIf authored144 items are MIT (github.com/TheoLeeCJ/SemIf at `ca3ba65f`), and
+the fixture carries their copyright and permission notice. The records written for the port use
+invented people, organisations and places.

SHA256SUMS

@@ -0,0 +1,31 @@
+bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a  LICENSE
+b0d697896e40d9cb44ab811247be0a9b22d99d8d79c50896818c3b42be65b058  README.md
+66f87f6fb2616b46604daf2a9c67ddc87938296d07156efa34d59b5be49e3238  config.json
+b9698210f9294274f55231d2ca521dcf3c28d767800508af89303f025c33ef76  gpu-pipelined/clef_flash_decode_fp16_pf64/clef_flash_decode_fp16_pf64.aimodel/main.hash
+0b7dfa30a0597ead745d8f3ed9ed11fa5ac639a4c578a33a7cca7ced11b9bea8  gpu-pipelined/clef_flash_decode_fp16_pf64/clef_flash_decode_fp16_pf64.aimodel/main.mlirb
+eab5bf891361488d40409bdf287f0c5379ee820f25d421d468e765508cbb8f50  gpu-pipelined/clef_flash_decode_fp16_pf64/clef_flash_decode_fp16_pf64.aimodel/metadata.json
+bbd4401e72dd055c9fb9998aecf3fcd1b43ea789698bec9040c0d89412a77ae3  gpu-pipelined/clef_flash_decode_fp16_pf64/metadata.json
+a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715  gpu-pipelined/clef_flash_decode_fp16_pf64/tokenizer/chat_template.jinja
+06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523  gpu-pipelined/clef_flash_decode_fp16_pf64/tokenizer/tokenizer.json
+91a08f825d370d085d692e04cf117cdd7faad7bf18e996f1e6031b6dab03db72  gpu-pipelined/clef_flash_decode_fp16_pf64/tokenizer/tokenizer_config.json
+79cfb36d716b67f5f945f5819ef1342b278959867a541f449a6a2e09b7af571b  gpu-pipelined/clef_flash_decode_int8mix_pf64/clef_flash_decode_int8mix_pf64.aimodel/main.hash
+2c19a0e765a659e0f726b4b7099a62490fed6ab2e4db1246eb63fd70996a67b9  gpu-pipelined/clef_flash_decode_int8mix_pf64/clef_flash_decode_int8mix_pf64.aimodel/main.mlirb
+4e838a8bd4ba48a1e5ca5cbd340a32d969ab8287984af4018aefde10372333e8  gpu-pipelined/clef_flash_decode_int8mix_pf64/clef_flash_decode_int8mix_pf64.aimodel/metadata.json
+1eafaedf47981c83c82bf4f609615e2a33f28093b8fab3d89f86dec0e520fc36  gpu-pipelined/clef_flash_decode_int8mix_pf64/metadata.json
+a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715  gpu-pipelined/clef_flash_decode_int8mix_pf64/tokenizer/chat_template.jinja
+06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523  gpu-pipelined/clef_flash_decode_int8mix_pf64/tokenizer/tokenizer.json
+91a08f825d370d085d692e04cf117cdd7faad7bf18e996f1e6031b6dab03db72  gpu-pipelined/clef_flash_decode_int8mix_pf64/tokenizer/tokenizer_config.json
+1d3470bf90972417bc5bad9e3930080a05206e887a36f15a1e70fe983812b806  gpu-pipelined/clef_flash_g256_vision_fp16w32/clef_flash_g256_vision_fp16w32.aimodel/main.hash
+b2adb0396f6f37f3178c0d7b413d10cddc31271c0dc7a56b2416125e9664f31d  gpu-pipelined/clef_flash_g256_vision_fp16w32/clef_flash_g256_vision_fp16w32.aimodel/main.mlirb
+630d7f7cbba089b019bd50db9ce9a4901232fca86415b4ff66ae1d9eb944f3bb  gpu-pipelined/clef_flash_g256_vision_fp16w32/clef_flash_g256_vision_fp16w32.aimodel/metadata.json
+4f01fd00bb557446ba3c64dec72dfe77f53313667c0c35fbc3186b6dacae6d06  gpu-pipelined/clef_flash_g448_vision_fp16w32/clef_flash_g448_vision_fp16w32.aimodel/main.hash
+04d8629bedeb1af76a2e63ed677a931c187289a6478fc66bf33a7bfa523a1eaa  gpu-pipelined/clef_flash_g448_vision_fp16w32/clef_flash_g448_vision_fp16w32.aimodel/main.mlirb
+b40a1fff26dbc2ba59d7583f8bf804b58c3d1d536e51abfbd7196fdd05ba2145  gpu-pipelined/clef_flash_g448_vision_fp16w32/clef_flash_g448_vision_fp16w32.aimodel/metadata.json
+f5fcb38a55e72369072fc1623a49bcd1b69209830abff7ac1c057955b7223a08  gpu-pipelined/clef_flash_head_bucket_fp16w32/clef_flash_head_bucket_fp16w32.aimodel/main.hash
+51e512ad1e4d3bcfdc7af43c1100124bd94d539de01f09651afc9b98486856fe  gpu-pipelined/clef_flash_head_bucket_fp16w32/clef_flash_head_bucket_fp16w32.aimodel/main.mlirb
+bfa78017346ae7bbf9e6709dff02a4c53ee2fd749d2307fe45e5c1c4ab50e4f7  gpu-pipelined/clef_flash_head_bucket_fp16w32/clef_flash_head_bucket_fp16w32.aimodel/metadata.json
+c7c4f1b1c054d938eb3523da52505a93eda8d253d94ab7cc20d2192724971fbe  gpu-pipelined/clef_flash_head_bucket_fp16w32/metadata.json
+77efe959a38b5b17b241543e129e695f3c77465ece55a25985279bd8176279a0  host/joint_head_config.json
+bfccf00b8d5d3f6d810839a52c2ec627700a9aaebaa26aedcb96931bd21ba36d  host/lm_head_fp16.bin
+ef89e2fb12d286cff9a4069365bfddad1a3455af0db7f17d7bc3156e4b75e86a  host/lm_head_fp16.json
+77efe959a38b5b17b241543e129e695f3c77465ece55a25985279bd8176279a0  joint_head_config.json

config.json

@@ -0,0 +1,109 @@
+{
+  "architectures": [
+    "Qwen3_5ForConditionalGeneration"
+  ],
+  "dtype": "bfloat16",
+  "image_token_id": 248056,
+  "model_type": "qwen3_5",
+  "text_config": {
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "attn_output_gate": true,
+    "bos_token_id": null,
+    "dtype": "bfloat16",
+    "eos_token_id": 248044,
+    "full_attention_interval": 4,
+    "head_dim": 256,
+    "hidden_act": "silu",
+    "hidden_size": 4096,
+    "initializer_range": 0.02,
+    "intermediate_size": 12288,
+    "layer_types": [
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention"
+    ],
+    "linear_conv_kernel_dim": 4,
+    "linear_key_head_dim": 128,
+    "linear_num_key_heads": 16,
+    "linear_num_value_heads": 32,
+    "linear_value_head_dim": 128,
+    "mamba_ssm_dtype": "float32",
+    "max_position_embeddings": 262144,
+    "mlp_only_layers": [],
+    "model_type": "qwen3_5_text",
+    "mtp_num_hidden_layers": 0,
+    "mtp_use_dedicated_embeddings": false,
+    "num_attention_heads": 16,
+    "num_hidden_layers": 32,
+    "num_key_value_heads": 4,
+    "pad_token_id": null,
+    "partial_rotary_factor": 0.25,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "mrope_interleaved": true,
+      "mrope_section": [
+        11,
+        11,
+        10
+      ],
+      "partial_rotary_factor": 0.25,
+      "rope_theta": 10000000,
+      "rope_type": "default"
+    },
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "vocab_size": 248320
+  },
+  "tie_word_embeddings": false,
+  "transformers_version": "5.10.2",
+  "video_token_id": 248057,
+  "vision_config": {
+    "deepstack_visual_indexes": [],
+    "depth": 27,
+    "dtype": "bfloat16",
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1152,
+    "in_channels": 3,
+    "initializer_range": 0.02,
+    "intermediate_size": 4304,
+    "model_type": "qwen3_5_vision",
+    "num_heads": 16,
+    "num_position_embeddings": 2304,
+    "out_hidden_size": 4096,
+    "patch_size": 16,
+    "spatial_merge_size": 2,
+    "temporal_patch_size": 2
+  },
+  "vision_end_token_id": 248054,
+  "vision_start_token_id": 248053
+}

gpu-pipelined/clef_flash_decode_fp16_pf64/clef_flash_decode_fp16_pf64.aimodel/main.hash

@@ -0,0 +1 @@
+}�0�Y~�t]�>���Z�9��x�:|�|����

gpu-pipelined/clef_flash_decode_fp16_pf64/clef_flash_decode_fp16_pf64.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b7dfa30a0597ead745d8f3ed9ed11fa5ac639a4c578a33a7cca7ced11b9bea8
+size 15880253283

gpu-pipelined/clef_flash_decode_fp16_pf64/clef_flash_decode_fp16_pf64.aimodel/metadata.json

@@ -0,0 +1,5 @@
+{
+  "assetVersion" : "2.0",
+  "producer" : "coreai-core 1.0.0b2",
+  "creationDate" : "20261003T025329Z"
+}

gpu-pipelined/clef_flash_decode_fp16_pf64/metadata.json

@@ -0,0 +1,148 @@
+{
+  "metadata_version": "0.2",
+  "kind": "decision-backbone",
+  "name": "clef_flash_decode_fp16_pf64",
+  "assets": {
+    "main": "clef_flash_decode_fp16_pf64.aimodel"
+  },
+  "language": {
+    "tokenizer": "Cloudflare/clef-flash",
+    "vocab_size": 248320,
+    "max_context_length": 4096,
+    "embedded_tokenizer": true,
+    "prefill_chunk": 64,
+    "static_inputs": [
+      "image_embeds",
+      "image_rc",
+      "rope_shift_start",
+      "rope_shift_amount"
+    ],
+    "image_tokens_max": 1024,
+    "output": "hidden [1, 64, 4096] fp16, every position",
+    "function_map": {
+      "main": [
+        "main"
+      ]
+    }
+  },
+  "source": {
+    "model_definition": "torch",
+    "hf_model_id": "Cloudflare/clef-flash",
+    "hf_revision": "17f0b0ad64efb65d273590632833508766b2aae6"
+  },
+  "compression": null,
+  "decision": {
+    "output": "hidden [1, S, 4096] per call: the final-norm hidden state at every position (no vocabulary head in the graph)",
+    "readout": "fresh zero states; a row of T ids runs as ceil(T / 64) calls of 'main' (static S = 64): call k gets ids[64k : 64k + 64] with position_ids 0..64k+63; the last call is padded with <|endoftext|> (248044) and the hidden rows of the padded positions are discarded (causal: they cannot reach a real position). The hidden rows [1, 64, 4096] of every call, concatenated and cut to T, are the backbone's final-norm last_hidden_state [T, 4096] the head reads.",
+    "prompt": {
+      "source": "the author's encode_record() (joint_schema_model.py at the pinned revision, sha256 0e304cf7c6500e8bb59bef7e2afd2c6373f82596dfb3b57d1aa93c175e2dc3a3); every piece tokenized on its own without special tokens, then concatenated (conversion/clef_flash/host.py build_ids)",
+      "layout": "prefix + [image block] + state + schema + suffix",
+      "prefix": "<|im_start|>system\\n{system prompt}<|im_end|>\\n<|im_start|>user\\nSTATE:\\n",
+      "prefix_tokens": 36,
+      "system_prompt": "Read the complete state and schema. Decide every field jointly. Each answer must be exactly one of that field's allowed options.",
+      "image_block": "at index 36 (right after the prefix): <|vision_start|> (248053), N <|image_pad|> (248056) sent to the graph as ids V+k (k = 0..N-1 row-major over the merged grid, V = 248320), <|vision_end|> (248054), '\\n'; N + 3 ids",
+      "state": "render(state): a string as is, anything else compact JSON (sorted keys, ensure_ascii off); cut so the whole row fits max_length",
+      "schema": "'\\n\\nSCHEMA FIELDS:\\n', then per question i (1-based) '\\nFIELD {i}\\nID: {id}\\nTYPE: {type}\\nINSTRUCTION: ' + render(instructions or id) + '\\nALLOWED OPTIONS:\\n' + per option j (1-based) 'OPTION {j}: ' + render({option_id, description}) + '\\n', then 'END FIELD\\n'",
+      "suffix": "\\n<|im_end|>\\n<|im_start|>assistant\\n<think>\\n\\n</think>\\n\\nJOINT SCHEMA DECISIONS:",
+      "suffix_tokens": 18,
+      "last_token": "':' (25)",
+      "add_special_tokens": false,
+      "chat_template": null,
+      "max_length_author": 16384,
+      "max_tokens_this_graph": "max_context_length (the row and its last padded call must fit)"
+    },
+    "spans": {
+      "question": "[start, end) of the render(instructions or id) tokens of the question",
+      "option": "[start, end) of the render({option_id, description}) tokens of each option",
+      "ids": "spans index the processor-form ids (N x <|image_pad|>), the same positions as the graph's"
+    },
+    "option_order": {
+      "noul": [
+        "true",
+        "false"
+      ],
+      "choice": "option ids sorted as strings",
+      "score": "levels 0..n-1"
+    },
+    "question_types": {
+      "noul": 0,
+      "choice": 1,
+      "score": 2
+    },
+    "noul_default_criteria": {
+      "true": "The proposition is true or the answer is yes.",
+      "false": "The proposition is false or the answer is no."
+    },
+    "head": {
+      "files": [
+        "joint_head.safetensors",
+        "joint_head_config.json"
+      ],
+      "code": "JointSchemaHead in the checkpoint's joint_schema_model.py (sha256 0e304cf7c6500e8bb59bef7e2afd2c6373f82596dfb3b57d1aa93c175e2dc3a3)",
+      "reads": "last_hidden_state [T, 4096] at every position; LayerNorm(4096) -> memory; the global vector = the last token; question / option vectors = the mean of the LayerNorm-ed rows over each span",
+      "lexical": "lm_head.weight rows (untied [248320, 4096], model-00001-of-00004.safetensors) averaged over each option span's ids; the table is not in this graph",
+      "dtype_of_record": "fp32 (the gate reads the graph's fp16 hidden through the fp32 head)"
+    },
+    "softmax": "per question, fp32, over that question's options (temperature 1)",
+    "response": {
+      "shape": "SystemOne-compatible: {model, answers: {question_id: answer}, usage: {input_tokens: T, output_tokens: 0}}",
+      "noul": "{type: 'noul', noul: p(true)}",
+      "choice": "{type: 'choice', choice: argmax option id, confidence: its p, probabilities: {id: p}}",
+      "score": "{type: 'score', score: sum(level * p), confidence: max p, legend: {level: criterion}, probabilities: {level: p}}",
+      "rounding": "4 decimals"
+    }
+  },
+  "vision": {
+    "n_image_max": 1024,
+    "image_token_base": 248320,
+    "image_embeds": "tower rows 0..N-1 (N = H*W merged tokens) as fp16, rows N.. zero",
+    "image_rc": "image_rc[k] = (k // W, k % W) for k < N, rows N.. zero",
+    "rope_shift_start": "37 + H*W (the <|vision_end|> index; the image block starts at 36)",
+    "rope_shift_amount": "H*W - max(H, W)",
+    "text_only": {
+      "rope_shift_start": 1073741824,
+      "rope_shift_amount": 0,
+      "image_embeds": "zero",
+      "image_rc": "zero"
+    },
+    "towers": [
+      {
+        "name": "clef_flash_g256_vision_fp16w32",
+        "merged_grid": [
+          8,
+          8
+        ],
+        "tile": 256,
+        "patches": [
+          256,
+          1536
+        ],
+        "image_embeds": [
+          64,
+          4096
+        ]
+      },
+      {
+        "name": "clef_flash_g448_vision_fp16w32",
+        "merged_grid": [
+          14,
+          14
+        ],
+        "tile": 448,
+        "patches": [
+          784,
+          1536
+        ],
+        "image_embeds": [
+          196,
+          4096
+        ]
+      }
+    ],
+    "host_preprocess": "RGB -> Pillow BICUBIC resize to (32*grid) x (32*grid), aspect not kept -> /255 -> (x - 0.5) / 0.5 -> merge-block-major patchify, the frame repeated at both temporal slots (conversion/clef_flash/host.py preprocess)"
+  },
+  "compilation": {
+    "date": "2026-10-03T02:53:35.119832+00:00",
+    "targets": []
+  }
+}

gpu-pipelined/clef_flash_decode_fp16_pf64/tokenizer/chat_template.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

gpu-pipelined/clef_flash_decode_fp16_pf64/tokenizer/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+size 19989325

gpu-pipelined/clef_flash_decode_fp16_pf64/tokenizer/tokenizer_config.json

@@ -0,0 +1,30 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}

gpu-pipelined/clef_flash_decode_int8mix_pf64/clef_flash_decode_int8mix_pf64.aimodel/main.hash

@@ -0,0 +1 @@
+,��e�Y��&��	�bI�j���F�c�p�jg�

gpu-pipelined/clef_flash_decode_int8mix_pf64/clef_flash_decode_int8mix_pf64.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c19a0e765a659e0f726b4b7099a62490fed6ab2e4db1246eb63fd70996a67b9
+size 11826520241

gpu-pipelined/clef_flash_decode_int8mix_pf64/clef_flash_decode_int8mix_pf64.aimodel/metadata.json

@@ -0,0 +1,5 @@
+{
+  "assetVersion" : "2.0",
+  "producer" : "coreai-core 1.0.0b2",
+  "creationDate" : "20261003T035712Z"
+}

gpu-pipelined/clef_flash_decode_int8mix_pf64/metadata.json

@@ -0,0 +1,272 @@
+{
+  "metadata_version": "0.2",
+  "kind": "decision-backbone",
+  "name": "clef_flash_decode_int8mix_pf64",
+  "assets": {
+    "main": "clef_flash_decode_int8mix_pf64.aimodel"
+  },
+  "language": {
+    "tokenizer": "Cloudflare/clef-flash",
+    "vocab_size": 248320,
+    "max_context_length": 4096,
+    "embedded_tokenizer": true,
+    "prefill_chunk": 64,
+    "static_inputs": [
+      "image_embeds",
+      "image_rc",
+      "rope_shift_start",
+      "rope_shift_amount"
+    ],
+    "image_tokens_max": 1024,
+    "output": "hidden [1, 64, 4096] fp16, every position",
+    "function_map": {
+      "main": [
+        "main"
+      ]
+    }
+  },
+  "source": {
+    "model_definition": "torch",
+    "hf_model_id": "Cloudflare/clef-flash",
+    "hf_revision": "17f0b0ad64efb65d273590632833508766b2aae6"
+  },
+  "compression": {
+    "scheme": "int8mix",
+    "linear": "int8 per-block-32 symmetric_with_clipping (weight only)",
+    "excluded": "fp16: the embedding table, the GDN conv1d, every RMSNorm (RMSNorm, RMSNormPlusOne, RMSNormGated), SDPA, RoPE, the GDN A_log / dt_bias",
+    "excluded_types": [
+      "coreai_models.primitives.macos.rms_norm.RMSNorm",
+      "coreai_models.primitives.macos.rms_norm.RMSNormGated",
+      "coreai_models.primitives.macos.rms_norm.RMSNormPlusOne",
+      "coreai_models.primitives.macos.rope.RoPE",
+      "coreai_models.primitives.macos.sdpa.SDPA",
+      "torch.nn.modules.conv.Conv1d",
+      "torch.nn.modules.sparse.Embedding"
+    ],
+    "int8_linear_modules": 155,
+    "fp16_layers": [
+      0,
+      1,
+      2,
+      3,
+      4,
+      5,
+      6,
+      7,
+      8,
+      9,
+      10,
+      11
+    ],
+    "fp16_linear_modules": [
+      "model.layers.0.linear_attn.in_proj_qkv",
+      "model.layers.0.linear_attn.in_proj_z",
+      "model.layers.0.linear_attn.in_proj_b",
+      "model.layers.0.linear_attn.in_proj_a",
+      "model.layers.0.linear_attn.out_proj",
+      "model.layers.0.mlp.gate_proj",
+      "model.layers.0.mlp.up_proj",
+      "model.layers.0.mlp.down_proj",
+      "model.layers.1.linear_attn.in_proj_qkv",
+      "model.layers.1.linear_attn.in_proj_z",
+      "model.layers.1.linear_attn.in_proj_b",
+      "model.layers.1.linear_attn.in_proj_a",
+      "model.layers.1.linear_attn.out_proj",
+      "model.layers.1.mlp.gate_proj",
+      "model.layers.1.mlp.up_proj",
+      "model.layers.1.mlp.down_proj",
+      "model.layers.2.linear_attn.in_proj_qkv",
+      "model.layers.2.linear_attn.in_proj_z",
+      "model.layers.2.linear_attn.in_proj_b",
+      "model.layers.2.linear_attn.in_proj_a",
+      "model.layers.2.linear_attn.out_proj",
+      "model.layers.2.mlp.gate_proj",
+      "model.layers.2.mlp.up_proj",
+      "model.layers.2.mlp.down_proj",
+      "model.layers.3.self_attn.q_proj",
+      "model.layers.3.self_attn.k_proj",
+      "model.layers.3.self_attn.v_proj",
+      "model.layers.3.self_attn.o_proj",
+      "model.layers.3.mlp.gate_proj",
+      "model.layers.3.mlp.up_proj",
+      "model.layers.3.mlp.down_proj",
+      "model.layers.4.linear_attn.in_proj_qkv",
+      "model.layers.4.linear_attn.in_proj_z",
+      "model.layers.4.linear_attn.in_proj_b",
+      "model.layers.4.linear_attn.in_proj_a",
+      "model.layers.4.linear_attn.out_proj",
+      "model.layers.4.mlp.gate_proj",
+      "model.layers.4.mlp.up_proj",
+      "model.layers.4.mlp.down_proj",
+      "model.layers.5.linear_attn.in_proj_qkv",
+      "model.layers.5.linear_attn.in_proj_z",
+      "model.layers.5.linear_attn.in_proj_b",
+      "model.layers.5.linear_attn.in_proj_a",
+      "model.layers.5.linear_attn.out_proj",
+      "model.layers.5.mlp.gate_proj",
+      "model.layers.5.mlp.up_proj",
+      "model.layers.5.mlp.down_proj",
+      "model.layers.6.linear_attn.in_proj_qkv",
+      "model.layers.6.linear_attn.in_proj_z",
+      "model.layers.6.linear_attn.in_proj_b",
+      "model.layers.6.linear_attn.in_proj_a",
+      "model.layers.6.linear_attn.out_proj",
+      "model.layers.6.mlp.gate_proj",
+      "model.layers.6.mlp.up_proj",
+      "model.layers.6.mlp.down_proj",
+      "model.layers.7.self_attn.q_proj",
+      "model.layers.7.self_attn.k_proj",
+      "model.layers.7.self_attn.v_proj",
+      "model.layers.7.self_attn.o_proj",
+      "model.layers.7.mlp.gate_proj",
+      "model.layers.7.mlp.up_proj",
+      "model.layers.7.mlp.down_proj",
+      "model.layers.8.linear_attn.in_proj_qkv",
+      "model.layers.8.linear_attn.in_proj_z",
+      "model.layers.8.linear_attn.in_proj_b",
+      "model.layers.8.linear_attn.in_proj_a",
+      "model.layers.8.linear_attn.out_proj",
+      "model.layers.8.mlp.gate_proj",
+      "model.layers.8.mlp.up_proj",
+      "model.layers.8.mlp.down_proj",
+      "model.layers.9.linear_attn.in_proj_qkv",
+      "model.layers.9.linear_attn.in_proj_z",
+      "model.layers.9.linear_attn.in_proj_b",
+      "model.layers.9.linear_attn.in_proj_a",
+      "model.layers.9.linear_attn.out_proj",
+      "model.layers.9.mlp.gate_proj",
+      "model.layers.9.mlp.up_proj",
+      "model.layers.9.mlp.down_proj",
+      "model.layers.10.linear_attn.in_proj_qkv",
+      "model.layers.10.linear_attn.in_proj_z",
+      "model.layers.10.linear_attn.in_proj_b",
+      "model.layers.10.linear_attn.in_proj_a",
+      "model.layers.10.linear_attn.out_proj",
+      "model.layers.10.mlp.gate_proj",
+      "model.layers.10.mlp.up_proj",
+      "model.layers.10.mlp.down_proj",
+      "model.layers.11.self_attn.q_proj",
+      "model.layers.11.self_attn.k_proj",
+      "model.layers.11.self_attn.v_proj",
+      "model.layers.11.self_attn.o_proj",
+      "model.layers.11.mlp.gate_proj",
+      "model.layers.11.mlp.up_proj",
+      "model.layers.11.mlp.down_proj"
+    ],
+    "head": "none in the graph (the untied lm_head table stays on the host)"
+  },
+  "decision": {
+    "output": "hidden [1, S, 4096] per call: the final-norm hidden state at every position (no vocabulary head in the graph)",
+    "readout": "fresh zero states; a row of T ids runs as ceil(T / 64) calls of 'main' (static S = 64): call k gets ids[64k : 64k + 64] with position_ids 0..64k+63; the last call is padded with <|endoftext|> (248044) and the hidden rows of the padded positions are discarded (causal: they cannot reach a real position). The hidden rows [1, 64, 4096] of every call, concatenated and cut to T, are the backbone's final-norm last_hidden_state [T, 4096] the head reads.",
+    "prompt": {
+      "source": "the author's encode_record() (joint_schema_model.py at the pinned revision, sha256 0e304cf7c6500e8bb59bef7e2afd2c6373f82596dfb3b57d1aa93c175e2dc3a3); every piece tokenized on its own without special tokens, then concatenated (conversion/clef_flash/host.py build_ids)",
+      "layout": "prefix + [image block] + state + schema + suffix",
+      "prefix": "<|im_start|>system\\n{system prompt}<|im_end|>\\n<|im_start|>user\\nSTATE:\\n",
+      "prefix_tokens": 36,
+      "system_prompt": "Read the complete state and schema. Decide every field jointly. Each answer must be exactly one of that field's allowed options.",
+      "image_block": "at index 36 (right after the prefix): <|vision_start|> (248053), N <|image_pad|> (248056) sent to the graph as ids V+k (k = 0..N-1 row-major over the merged grid, V = 248320), <|vision_end|> (248054), '\\n'; N + 3 ids",
+      "state": "render(state): a string as is, anything else compact JSON (sorted keys, ensure_ascii off); cut so the whole row fits max_length",
+      "schema": "'\\n\\nSCHEMA FIELDS:\\n', then per question i (1-based) '\\nFIELD {i}\\nID: {id}\\nTYPE: {type}\\nINSTRUCTION: ' + render(instructions or id) + '\\nALLOWED OPTIONS:\\n' + per option j (1-based) 'OPTION {j}: ' + render({option_id, description}) + '\\n', then 'END FIELD\\n'",
+      "suffix": "\\n<|im_end|>\\n<|im_start|>assistant\\n<think>\\n\\n</think>\\n\\nJOINT SCHEMA DECISIONS:",
+      "suffix_tokens": 18,
+      "last_token": "':' (25)",
+      "add_special_tokens": false,
+      "chat_template": null,
+      "max_length_author": 16384,
+      "max_tokens_this_graph": "max_context_length (the row and its last padded call must fit)"
+    },
+    "spans": {
+      "question": "[start, end) of the render(instructions or id) tokens of the question",
+      "option": "[start, end) of the render({option_id, description}) tokens of each option",
+      "ids": "spans index the processor-form ids (N x <|image_pad|>), the same positions as the graph's"
+    },
+    "option_order": {
+      "noul": [
+        "true",
+        "false"
+      ],
+      "choice": "option ids sorted as strings",
+      "score": "levels 0..n-1"
+    },
+    "question_types": {
+      "noul": 0,
+      "choice": 1,
+      "score": 2
+    },
+    "noul_default_criteria": {
+      "true": "The proposition is true or the answer is yes.",
+      "false": "The proposition is false or the answer is no."
+    },
+    "head": {
+      "files": [
+        "joint_head.safetensors",
+        "joint_head_config.json"
+      ],
+      "code": "JointSchemaHead in the checkpoint's joint_schema_model.py (sha256 0e304cf7c6500e8bb59bef7e2afd2c6373f82596dfb3b57d1aa93c175e2dc3a3)",
+      "reads": "last_hidden_state [T, 4096] at every position; LayerNorm(4096) -> memory; the global vector = the last token; question / option vectors = the mean of the LayerNorm-ed rows over each span",
+      "lexical": "lm_head.weight rows (untied [248320, 4096], model-00001-of-00004.safetensors) averaged over each option span's ids; the table is not in this graph",
+      "dtype_of_record": "fp32 (the gate reads the graph's fp16 hidden through the fp32 head)"
+    },
+    "softmax": "per question, fp32, over that question's options (temperature 1)",
+    "response": {
+      "shape": "SystemOne-compatible: {model, answers: {question_id: answer}, usage: {input_tokens: T, output_tokens: 0}}",
+      "noul": "{type: 'noul', noul: p(true)}",
+      "choice": "{type: 'choice', choice: argmax option id, confidence: its p, probabilities: {id: p}}",
+      "score": "{type: 'score', score: sum(level * p), confidence: max p, legend: {level: criterion}, probabilities: {level: p}}",
+      "rounding": "4 decimals"
+    }
+  },
+  "vision": {
+    "n_image_max": 1024,
+    "image_token_base": 248320,
+    "image_embeds": "tower rows 0..N-1 (N = H*W merged tokens) as fp16, rows N.. zero",
+    "image_rc": "image_rc[k] = (k // W, k % W) for k < N, rows N.. zero",
+    "rope_shift_start": "37 + H*W (the <|vision_end|> index; the image block starts at 36)",
+    "rope_shift_amount": "H*W - max(H, W)",
+    "text_only": {
+      "rope_shift_start": 1073741824,
+      "rope_shift_amount": 0,
+      "image_embeds": "zero",
+      "image_rc": "zero"
+    },
+    "towers": [
+      {
+        "name": "clef_flash_g256_vision_fp16w32",
+        "merged_grid": [
+          8,
+          8
+        ],
+        "tile": 256,
+        "patches": [
+          256,
+          1536
+        ],
+        "image_embeds": [
+          64,
+          4096
+        ]
+      },
+      {
+        "name": "clef_flash_g448_vision_fp16w32",
+        "merged_grid": [
+          14,
+          14
+        ],
+        "tile": 448,
+        "patches": [
+          784,
+          1536
+        ],
+        "image_embeds": [
+          196,
+          4096
+        ]
+      }
+    ],
+    "host_preprocess": "RGB -> Pillow BICUBIC resize to (32*grid) x (32*grid), aspect not kept -> /255 -> (x - 0.5) / 0.5 -> merge-block-major patchify, the frame repeated at both temporal slots (conversion/clef_flash/host.py preprocess)"
+  },
+  "compilation": {
+    "date": "2026-10-03T03:57:16.550932+00:00",
+    "targets": []
+  }
+}

gpu-pipelined/clef_flash_decode_int8mix_pf64/tokenizer/chat_template.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

gpu-pipelined/clef_flash_decode_int8mix_pf64/tokenizer/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+size 19989325

gpu-pipelined/clef_flash_decode_int8mix_pf64/tokenizer/tokenizer_config.json

@@ -0,0 +1,30 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}

gpu-pipelined/clef_flash_g256_vision_fp16w32/clef_flash_g256_vision_fp16w32.aimodel/main.hash

@@ -0,0 +1 @@
+���9oo7��

gpu-pipelined/clef_flash_g256_vision_fp16w32/clef_flash_g256_vision_fp16w32.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2adb0396f6f37f3178c0d7b413d10cddc31271c0dc7a56b2416125e9664f31d
+size 909207876

gpu-pipelined/clef_flash_g256_vision_fp16w32/clef_flash_g256_vision_fp16w32.aimodel/metadata.json

@@ -0,0 +1,5 @@
+{
+  "assetVersion" : "2.0",
+  "creationDate" : "20261003T010220Z",
+  "producer" : "coreai-core 1.0.0b2"
+}

gpu-pipelined/clef_flash_g448_vision_fp16w32/clef_flash_g448_vision_fp16w32.aimodel/main.hash

@@ -0,0 +1 @@
+�b����j.c�gz�r��G��k�:{�R:�

gpu-pipelined/clef_flash_g448_vision_fp16w32/clef_flash_g448_vision_fp16w32.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04d8629bedeb1af76a2e63ed677a931c187289a6478fc66bf33a7bfa523a1eaa
+size 911945013

gpu-pipelined/clef_flash_g448_vision_fp16w32/clef_flash_g448_vision_fp16w32.aimodel/metadata.json

@@ -0,0 +1,5 @@
+{
+  "assetVersion" : "2.0",
+  "producer" : "coreai-core 1.0.0b2",
+  "creationDate" : "20261003T010241Z"
+}

gpu-pipelined/clef_flash_head_bucket_fp16w32/clef_flash_head_bucket_fp16w32.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51e512ad1e4d3bcfdc7af43c1100124bd94d539de01f09651afc9b98486856fe
+size 243978965

gpu-pipelined/clef_flash_head_bucket_fp16w32/clef_flash_head_bucket_fp16w32.aimodel/metadata.json

@@ -0,0 +1,5 @@
+{
+  "creationDate" : "20261003T045330Z",
+  "assetVersion" : "2.0",
+  "producer" : "coreai-core 1.0.0b2"
+}

gpu-pipelined/clef_flash_head_bucket_fp16w32/metadata.json

@@ -0,0 +1,115 @@
+{
+  "metadata_version": "0.2",
+  "kind": "decision-head",
+  "name": "clef_flash_head_bucket_fp16w32",
+  "assets": {
+    "main": "clef_flash_head_bucket_fp16w32.aimodel"
+  },
+  "head": {
+    "graph": "conversion/clef_flash/clef_head.py ClefHeadGraph: the author's JointSchemaHead for one record, tensor inputs only",
+    "shape": "bucket",
+    "functions": [
+      "t512",
+      "t1024",
+      "t2048",
+      "t4096"
+    ],
+    "buckets": [
+      [
+        "t512",
+        512
+      ],
+      [
+        "t1024",
+        1024
+      ],
+      [
+        "t2048",
+        2048
+      ],
+      [
+        "t4096",
+        4096
+      ]
+    ],
+    "dims": null,
+    "host_pad": "bucket",
+    "host_pad_rule": "T to the smallest bucket >= T (512, 1024, 2048, 4096), Q to 16, O to 128; padding rows / columns zero, their valid flags 0",
+    "inputs": {
+      "hidden": "[T, 4096] f32",
+      "key_valid": "[T] f32",
+      "q_avg": "[Q, T] f32",
+      "o_avg": "[O, T] f32",
+      "g_avg": "[1, T] f32",
+      "lexical": "[O, 4096] f32",
+      "member": "[Q, O] f32",
+      "type_ids": "[Q] i32",
+      "q_valid": "[Q] f32",
+      "o_valid": "[O] f32"
+    },
+    "input_order": [
+      "hidden",
+      "key_valid",
+      "q_avg",
+      "o_avg",
+      "g_avg",
+      "lexical",
+      "member",
+      "type_ids",
+      "q_valid",
+      "o_valid"
+    ],
+    "outputs": {
+      "logits": "[O] f32; padding options 0"
+    },
+    "weights": "fp16w32",
+    "softmax": "host: per question, fp32, over that question's options",
+    "lexical": "host: lm_head_fp16.bin rows of each option span's ids, averaged (clef_head.lexical_rows)",
+    "source": {
+      "hf_id": "Cloudflare/clef-flash",
+      "revision": "17f0b0ad64efb65d273590632833508766b2aae6",
+      "joint_schema_model_sha256": "0e304cf7c6500e8bb59bef7e2afd2c6373f82596dfb3b57d1aa93c175e2dc3a3",
+      "joint_head_safetensors_sha256": "19cdcec8c81dc9212be320fff47462ab342fbc1278be4368fb3da71241cf5ba0",
+      "joint_head_config_sha256": "77efe959a38b5b17b241543e129e695f3c77465ece55a25985279bd8176279a0"
+    }
+  },
+  "export": {
+    "aimodel_bytes": 243979102,
+    "main_mlirb": {
+      "bytes": 243978965,
+      "sha256": "51e512ad1e4d3bcfdc7af43c1100124bd94d539de01f09651afc9b98486856fe"
+    },
+    "seconds": {
+      "build": 0.19440945898531936,
+      "export": 10.82508116599638,
+      "optimize": 0.031602208997355774,
+      "save": 0.2501100410008803,
+      "total": 11.301202874979936
+    },
+    "export_attempts": [
+      {
+        "externalize": "default externalize specs",
+        "ok": true,
+        "seconds": 9.44808970799204
+      }
+    ]
+  },
+  "aot": {
+    "note": "compiled for the gates with: xcrun coreai-build compile <this .aimodel> --output <dir> --platform macOS --preferred-compute gpu --architecture h16c (not shipped; the Swift host specializes the .aimodel)",
+    "flags": [
+      "--platform",
+      "macOS",
+      "--preferred-compute",
+      "gpu",
+      "--architecture",
+      "h16c"
+    ],
+    "seconds": 0.5527557500172406
+  },
+  "compilation": {
+    "date": "2026-10-03T13:53:31+09:00",
+    "targets": [
+      "h16c"
+    ]
+  }
+}

host/joint_head_config.json

@@ -0,0 +1,8 @@
+{
+  "hidden_size": 4096,
+  "width": 1024,
+  "routing_layers": 2,
+  "layers": 4,
+  "heads": 16,
+  "feedforward": 4096
+}

host/lm_head_fp16.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfccf00b8d5d3f6d810839a52c2ec627700a9aaebaa26aedcb96931bd21ba36d
+size 2034237440

host/lm_head_fp16.json

@@ -0,0 +1,88 @@
+{
+ "file": "lm_head_fp16.bin",
+ "shape": [
+  248320,
+  4096
+ ],
+ "dtype": "float16",
+ "byte_order": "little-endian",
+ "layout": "row-major, no header: row t at byte offset t * 8192",
+ "bytes": 2034237440,
+ "sha256": "bfccf00b8d5d3f6d810839a52c2ec627700a9aaebaa26aedcb96931bd21ba36d",
+ "row_t": "lm_head.weight[t] (the untied output embedding) rounded bf16 -> fp16, round to nearest even",
+ "use": "lexical[o] = mean over option o's span ids of the rows, accumulated in f32 or wider (clef_head.lexical_rows: f64); the head graph's `lexical` input",
+ "source": {
+  "hf_id": "Cloudflare/clef-flash",
+  "revision": "17f0b0ad64efb65d273590632833508766b2aae6",
+  "shard": "model-00001-of-00004.safetensors",
+  "shard_sha256": "8b45a8e968141cdcc58fb71c9adfc258e2c77b5f062bc636c1fd5bc5d916b565",
+  "tensor": "lm_head.weight",
+  "header": {
+   "dtype": "BF16",
+   "shape": [
+    248320,
+    4096
+   ],
+   "data_offsets": [
+    0,
+    2034237440
+   ]
+  },
+  "data_start": 4680
+ },
+ "rounding": {
+  "nonzero_elements": 1017118720,
+  "elements_below_fp16_normal_min": 3317193,
+  "elements_flushed_to_zero": 1649,
+  "elements_fp16_subnormal": 3315544,
+  "elements_inf": 0,
+  "max_rel_error_nonzero": 1.0,
+  "max_rel_error_not_flushed": 0.9844961166381836,
+  "max_rel_error_fp16_normal_range": 0.0,
+  "max_abs_error": 2.9802322387695312e-08,
+  "absmax": 0.35546875,
+  "fp16_normal_min": 6.103515625e-05,
+  "note": "bf16 values with |x| >= 2^-14 are exact in fp16; below that they become fp16 subnormals (rounded) or 0; max_rel_error_nonzero = 1.0 comes from the flushed ones"
+ },
+ "round1_check": {
+  "file": "results/lm_head_table.json",
+  "expected": {
+   "elements_below_fp16_normal_min": 3317193,
+   "elements_flushed_to_zero": 1649
+  },
+  "got": {
+   "elements_below_fp16_normal_min": 3317193,
+   "elements_flushed_to_zero": 1649
+  },
+  "equal": true,
+  "effect_at_the_head": "results/lm_head_table_effect.json: fp16 table max |dp| 1.8e-7 (= the fp32 re-run floor), argmax 405/405"
+ },
+ "readback": {
+  "rows_checked": 1200,
+  "first_rows": 100,
+  "last_rows": 100,
+  "random_rows": 1000,
+  "seed": 0,
+  "path": "safetensors.safe_open(framework='pt') get_slice -> torch bf16 -> float16",
+  "mismatched_elements": 0,
+  "bit_equal": true
+ },
+ "head_files": {
+  "joint_head.safetensors": {
+   "bytes": 243538016,
+   "sha256": "19cdcec8c81dc9212be320fff47462ab342fbc1278be4368fb3da71241cf5ba0",
+   "source": "Cloudflare/clef-flash@17f0b0ad64efb65d273590632833508766b2aae6: joint_head.safetensors"
+  },
+  "joint_head_config.json": {
+   "bytes": 119,
+   "sha256": "77efe959a38b5b17b241543e129e695f3c77465ece55a25985279bd8176279a0",
+   "source": "Cloudflare/clef-flash@17f0b0ad64efb65d273590632833508766b2aae6: joint_head_config.json"
+  }
+ },
+ "script_sha256": "663b7f5c65faa14b3387ae775b2d7e0f4cc7f3213466916de5fb11b7d00f7fb8",
+ "seconds": {
+  "convert_write": 5.24923904199386,
+  "total": 8.231139292009175
+ },
+ "written": "2026-10-03T13:41:43+09:00"
+}

joint_head_config.json

@@ -0,0 +1,8 @@
+{
+  "hidden_size": 4096,
+  "width": 1024,
+  "routing_layers": 2,
+  "layers": 4,
+  "heads": 16,
+  "feedforward": 4096
+}