Kev-4B Core AI: fp16 hidden-state decoder, Metal-kernel graph, 128 tokens per call (K128), with the pointer head and tokenizer

.gitattributes

@@ -33,3 +33,5 @@ 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/kev_4b_decode_fp16_metal_pf128/kev_4b_decode_fp16_metal_pf128.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
+gpu-pipelined/kev_4b_decode_fp16_metal_pf128/tokenizer/tokenizer.json 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 +1,373 @@
+---
+license: apache-2.0
+base_model: jaredpalmer/kev-4b
+base_model_relation: quantized
+pipeline_tag: text-classification
+library_name: coreai
+language:
+- en
+tags:
+- coreai
+- core-ai
+- apple
+- on-device
+- decision-model
+- systemone
+- typed-decisions
+- structured-output
+---
+
+# Kev-4B — Core AI
+
+Apple Core AI (`.aimodel`) conversion of [jaredpalmer/kev-4b](https://huggingface.co/jaredpalmer/kev-4b/tree/6cfce5c2fa4b4bd64026336ab649c5ca78857d52) (tag `v1.0`, Apache-2.0): a state and typed questions in, the probability of every option of every question out.
+
+| Variant | Path | Size | Requires | Tested on |
+|---|---|---:|---|---|
+| Decoder, fp16, with the host's pointer head | `gpu-pipelined/kev_4b_decode_fp16_metal_pf128/` | 8,431 MB | macOS 27 | M4 Max, macOS 27.0 (26A428), 2026-10-04 |
+
+One question of 94 tokens takes 100.2 ms on the M4 Max, in Swift ([measured below](#time-per-decision-on-the-mac)).
+
+```swift
+import Kev
+
+let bundle = URL(filePath: "Kev-4B-CoreAI/gpu-pipelined/kev_4b_decode_fp16_metal_pf128")
+let kev = try await KevDecider(bundle: bundle)              // asset: nil = the .aimodel, specialized here (GPU, frequent reshapes)
+let response = try await kev.decide(requestJSON: requestData, shared: true)   // shared: the state's whole calls run once
+print(PythonFormat.dumps(response, asciiOnly: false))
+```
+
+The snippet uses the [`Kev`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/Kev) 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/kev-4b/README.md); the rest of this page is the same text with repository links.
+
+Source [jaredpalmer/kev-4b](https://huggingface.co/jaredpalmer/kev-4b/tree/6cfce5c2fa4b4bd64026336ab649c5ca78857d52) (tag `v1.0`, commit `6cfce5c`) · base [Qwen/Qwen3.5-4B-Base](https://huggingface.co/Qwen/Qwen3.5-4B-Base/tree/1001bb4d826a52d1f399e183466143f4da7b741b) (revision `1001bb4`)
+
+A **decision model**. Give it a state (text, or a JSON object or array) and typed questions: `noul`
+(yes / no), `choice` (named options) or `score` (ordered levels). It returns a probability for every
+option of every question. 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.
+
+Jared Palmer trained it as a rank-16 LoRA adapter and a pointer head on Qwen3.5-4B-Base (32 layers:
+24 Gated DeltaNet and 8 full attention, hidden size 2,560). The author's card: "It is intended for
+developers who classify, route, triage or check documents and who need probabilities that can be
+thresholded, for example to send uncertain cases to human review." Each question is its own row: the
+state, then the question and its options. The head scores each option's closing token against the
+question's last token. The author's card reports benchmark results; none of them are re-measured here.
+
+This port merges the adapter into the base with the author's own merge script and exports the text
+backbone as one Core AI graph. The graph takes 128 token ids per call and returns the final-norm hidden
+state at every position; it has no vocabulary head. Each Gated DeltaNet layer runs its recurrence in an
+fp32 Metal kernel, one GPU dispatch per layer per call. Weights are fp16 (8.41 GB). The host runs the
+pointer head from the author's head weights. The gate is probability parity with the author's own fp32
+code, on every option of every question, on a fixture of 434 questions and on a held-out set of 130.
+
+It runs on the Mac GPU. This release is for the Mac: the one load of an iPhone AOT asset of an earlier
+graph on an iPhone 18 Pro crashed (below). [Kev-0.8B](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/README.md), the same port at 0.8B, is
+measured on the iPhone 18 Pro.
+
+## Readout contract
+
+The contract is Kev-0.8B's, with this checkpoint's head and temperature; the full description is in the
+[Kev-0.8B card](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/README.md#readout-contract). In short:
+
+```
+[<|fim_prefix|>] + render(state) + [<|fim_middle|>] + render(instructions)
+  + for each option: [<|box_start|>] + option_text + [<|box_end|>]
+  + [<|fim_suffix|>]
+```
+
+- One row per question, zeroed states, positions 0..L − 1. Delimiters from the base tokenizer: 248060,
+  248061, 248049 / 248050, 248062; pad = eos = 248044; no bos. The 4B base's `tokenizer.json` is
+  byte-identical to the 0.8B base's; its `tokenizer_config.json` differs in the chat template's thinking
+  default, which this readout does not use. The `kev-4b` repository still carries Qwen3-era
+  `vocab.json`, `merges.txt` and `added_tokens.json` (other delimiter ids); the author's loader reads the
+  tokenizer from the base, and so does this port.
+- The head: `z_k = ((W_k h_opt_k + b_k) · (W_q h_decide + b_q)) / 16`, `p = softmax(z / T)` within the
+  question, T = 2.406050072164233 (`head.pt`); `W_q`, `W_k` are `[256, 2560]` with biases, fp32.
+- A row holds at most **3,968 tokens**. The last call is padded to 128 tokens and must end at or below
+  the graph's position bound of 4,095.
+- Shared prefix (optional, exact): the state's whole 128-token calls run once and the four states are
+  copied per question. Every row's hidden state equals its direct run bit for bit (below).
+
+The fixture and the held-out set are Kev-0.8B's (434 and 130 questions), with this checkpoint's own fp32
+oracle run (the author's package, one CPU thread): [`fixtures-kev-4b.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/fixtures-kev-4b.json). It has
+twelve near-ties on the fixture and four on the held-out set.
+
+## Core AI shape
+
+The same module as Kev-0.8B, `conversion/kev/qwen3_5_kev_decoder.py`, with no code change: the overlay's
+stateful Qwen3.5 text decoder on the merged weights, no vocabulary head. One function, `main`, at a
+static 128 tokens. Each Gated DeltaNet layer runs its recurrence in the overlay's fp32 Metal chunk kernel
+(`qwen3_5_gdn_metal`): one GPU dispatch per layer per call, the recurrent state kept in fp32 through the
+call.
+
+| | name | shape, type |
+|---|---|---|
+| inputs | `input_ids` | [1, 128] int32 |
+| | `position_ids` | [1, seq] int32, the ramp 0 .. seq − 1 |
+| 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, 128, 2560] fp16, every position |
+
+Per row of T ids, from zeroed states: ⌈T / 128⌉ calls, the last padded with `<|endoftext|>` (248044) and
+its padded rows dropped. `metadata.json` says `kind: decision-backbone` and `language.prefill_chunk: 128`,
+and carries the readout contract. The host (`conversion/kev/decide.py`, [`apps/Kev`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/Kev)) is
+Kev-0.8B's: the head in float64 from `head/head.safetensors`, p rounded to fp32 once, the
+SystemOne-compatible response, the shared prefix and the prepared state. No engine, no runtime patch.
+
+## Measured (Apple M4 Max, macOS 27.0 26A428, 2026-10-03 – 10-04)
+
+The bar is Kev-0.8B's, 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 rows of each row's mean |Δp| ≤ 0.002, every process's re-run bit-equal.
+The Python gates load the AOT `.aimodelc` (h16c, `--expect-frequent-reshapes`) with
+`SpecializationOptions.default()`; their times are on a GPU shared with other work.
+
+### Before any graph: the merge and fp32 torch
+
+- The merged checkpoint (`scripts/merge_lora_checkpoint.py`, tag `kev-1.0`; fp32, four shards of
+  16.8 GB, 426 tensors, 248 of them adapted) answers through the author's loader as the adapter
+  checkpoint does: 434/434 questions, logits bit-equal. `W + (B @ A) × 2` reproduces three merged
+  tensors bit for bit (the adapter moves them by 1.4–2.0 %)
+  ([`gate-kev-4b-merge.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/gate-kev-4b-merge.json)).
+- The decoder module in fp32 on the CPU, driven like the graph with the recurrence unrolled, read through
+  the author's head: 434/434 argmax (the 12 near-ties included), max |Δp| 6.6e-6, lowest per-position
+  hidden cosine 0.9999999965 on 21 rows; the module's loader reads all 426 keys and leaves none; it equals
+  the overlay's plain text decoder bit for bit (3 rows, one of 1,518 tokens)
+  ([`gate-kev-4b-torch-parity.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/gate-kev-4b-torch-parity.json)).
+- The Metal kernel has no torch values: its `torch_defn` returns zeros of the right shapes. The GPU gate
+  below is its parity check.
+
+### The graph alone on the Mac GPU
+
+| set | questions | argmax (margin > 0.02) | near-ties agreeing | max \|Δp\| | mean of row means | bar |
+|---|---:|---:|---:|---:|---:|---|
+| **fixture** | 434 | **422/422** | **10/12** | **0.0153** | **0.00077** | **PASS** |
+| held out | 130 | 126/126 | 3/4 | 0.0096 | 0.00069 | PASS |
+
+The near-ties that flip have oracle margins of 0.0005–0.0024 and move by at most 0.0029. The worst
+fixture row is a SemIf item (0.0153), the worst held-out row a `score` item (0.0096). The lowest
+per-position hidden cosine on the 21 recorded rows is 0.9921. Every value is finite and every process's
+re-run is bit-equal. The red arms move p: a swapped state moves 5 of 5 argmaxes (max |Δp| 0.988), a
+grammatical "not" 3 of 5 (0.957). Transcript: [`gate-kev-4b-readout.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/gate-kev-4b-readout.json).
+
+### From the request: Python and Swift
+
+- `conversion/kev/host.py`, without the author's package, rebuilds every oracle row from the raw request
+  (ids, `<|fim_suffix|>` / `<|box_end|>` indices, keys): 434/434 and 130/130 rows with two tokenizer
+  implementations. `decide.py` on the AOT graph: hidden rows bit-equal to the gate's on 434/434 and
+  130/130 rows, and the shared prefix equal to the direct run on every row
+  ([`gate-kev-4b-host.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/gate-kev-4b-host.json)).
+- The Swift host ([`apps/Kev`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/Kev), Release): the ids from the raw request equal the
+  oracle's on 434/434 and 130/130 rows. On the same AOT asset every row's hidden state equals the Python
+  gate's bit for bit (434/434 and 130/130), and its float64 head gives p within 4.5e-7 of the gate's fp32
+  head. The bar is the gate's (fixture 0.0153 / 0.00077, held out 0.0096 / 0.00069). With the shared
+  prefix every row's hidden state and p equal the direct run's (434/434 and 130/130).
+- **JIT = AOT.** The `.aimodel` specialized by the Swift runtime (GPU preferred,
+  `expectFrequentReshapes`) gives the AOT asset's hidden state and p bit for bit on 25 rows of 10 records.
+  The specialization took 16.6 s in `AIModel(contentsOf:)` (18.4 s with the tokenizer and the head)
+  and left a 15.55 GB entry in the runtime's cache; a later load took 1.74 s
+  ([`gate-kev-4b-swift.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/gate-kev-4b-swift.json)).
+
+### Time per decision on the Mac
+
+Swift, Release CLI, the AOT asset, in one machine-wide GPU lock window (2026-10-04 14:05–14:30 JST). Two
+processes per graph counted only when no other GPU job ran and the one-minute load average was at most 12;
+each made 6 decisions per item after one warm-up, and the table gives the median (p10–p90) over the 12.
+The earlier unrolled graph (16 tokens per call) ran in the same window:
+
+| request | tokens | calls | ms | p10–p90 | unrolled S = 16, ms |
+|---|---:|---:|---:|---|---:|
+| one question | 94 | 1 | 100.2 | 99.3–101.2 | 269.5 |
+| one question | 380 | 3 | 298.0 | 296.1–304.5 | 1,075.4 |
+| one question | 1,518 | 12 | 1,189.0 | 1,182.9–1,194.9 | 4,341.9 |
+| one question | 1,802 | 15 | 1,488.1 | 1,484.3–1,498.5 | 5,200.2 |
+| five questions on one 137-token state, each row from zero | 240 | 10 | 986.7 | 983.0–990.2 | 2,367.9 |
+| the same five questions, the state run once (shared) | 240 | 6 | 615.6 | 613.3–624.5 | 894.7 |
+| eight questions on that state, each row from zero | 320 | 16 | 1,578.7 | 1,574.2–1,591.6 | 3,776.0 |
+| the same eight questions, shared | 320 | 9 | 920.0 | 916.9–930.6 | 1,262.4 |
+| four questions on one 1,477-token state, each row from zero | 1,622 | 49 | 4,851.7 | 4,839.2–4,870.6 | 17,087.7 |
+| the same four questions, shared | 1,622 | 16 | 1,614.5 | 1,607.5–1,647.8 | 4,713.7 |
+
+Loading the AOT asset took 9.90 / 9.91 s at the start of each process and 1.30 / 1.22 s when loaded
+again. The two processes ended at 0.89 and 0.95 GB of footprint. Transcript:
+[`gate-kev-4b-timing-mac.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/gate-kev-4b-timing-mac.json).
+
+### iPhone 18 Pro: not in this release
+
+An earlier graph of this port (the recurrence unrolled, 16 tokens per call) compiled for the iPhone 18
+Pro's GPU (h19p, `--expect-frequent-reshapes`) is 15,557,759,365 bytes. Loaded once in the headless gate
+app (iOS 27.0 24A437, `.default` options, no increased-memory-limit entitlement), the app died inside
+`AIModel(contentsOf:)` with `EXC_BAD_ACCESS (SIGSEGV)` in the on-device compile for delegates
+(`-[MPSGraphAICodeCompilerDelegate getInitializedAICodeBytecodeWithPayloadPrefix:delegateId:]`). Its last
+memory sample, 0.32 s in, read a footprint of 175.9 MB with 3,364.1 MB available. The cause is not
+isolated and the load was not retried; the graph of this release was not loaded on the phone
+([`../kev-0.8b/gate-kev-0.8b-iphone.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/gate-kev-0.8b-iphone.json)
+`round8.device.summary.load_4b`, `round8.load_4b_memory_samples`).
+
+## Forms measured
+
+Every form below was exported and compiled the same way and read against the same oracle: the full gate,
+or the 130-row subset where the table says so. The times come from different windows, so compare forms only
+within one window. Full records: [`gate-kev-4b-forms.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/gate-kev-4b-forms.json).
+
+Mac (Swift Release CLI, the 94-token question, median of the counted processes):
+
+| form | 94-token decision, ms | window | fixture max \|Δp\| | note |
+|---|---:|---|---:|---|
+| Gated DeltaNet unrolled, S = 16 | 269.5 | round 12 | 0.0154 | |
+| unrolled, S = 32 / 64 | — | — | 0.0078 / 0.0070 (130 rows) | not timed in a lock window (104.8 / 174.1 ms per call on a shared GPU) |
+| Metal kernel, S = 32 | — | — | 0.0074 (130 rows) | not timed in a lock window (48.4 ms per call on a shared GPU) |
+| Metal kernel, S = 64 | 118.0 | round 12 | 0.0170 | five questions shared: 435.9 ms (S = 128: 615.6) |
+| **Metal kernel, S = 128 (this release)** | **100.2** | round 12 | **0.0153** | |
+| Metal kernel, query length 2..512, host calls of at most 512 ids | 83.3 | round 12 | 0.0132 | not shipped: memory grows (below) |
+| in-graph chunk scan | — | — | — | not run at 4B; on Kev-0.8B it fails from S = 32 |
+| int8, every linear (unrolled S = 16) | — | — | 0.0601 | fails the bar (Precision) |
+| any form on the Neural Engine | — | — | — | not run: the recurrence's fp32 state is not an ANE element type ([`qwen3.5-static-ane.md`](https://github.com/john-rocky/coreai-model-zoo/blob/main/knowledge/qwen3.5-static-ane.md)) |
+
+**S = 64 for several short questions on one state.** In the same window, five questions on one
+137-token state, shared, took 435.9 ms with S = 64 and 615.6 ms with S = 128; one 94-token question took
+118.0 and 100.2 ms. This release ships S = 128.
+
+**Why the dynamic-length graph does not ship.** It passed the gate and decided the 94-token question in
+83.3 ms in the same window. Its two timing processes ended at 6.31 and 6.32 GB of footprint, against 0.89
+and 0.95 GB for S = 128. On Kev-0.8B the same form keeps growing while the call length keeps changing,
+until the `AIModel` is created again ([Kev-0.8B card](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/README.md#forms-measured)).
+
+## Precision
+
+**int8 was measured and is not shipped.** It was measured on the unrolled S = 16 graph; the Metal-kernel
+graphs were not measured in int8. int8 per block of 32 over every decoder linear fails the bar:
+
+| decoder (unrolled S = 16) | fp16 kept | `main.mlirb` bytes | fixture max \|Δp\| / mean | held out max \|Δp\| / mean | bar |
+|---|---:|---:|---|---|---|
+| fp16 | 100 % | 8,414,007,636 | 0.0154 / 0.00078 | 0.0109 / 0.00072 | PASS |
+| int8lin: every linear int8 (`symmetric_with_clipping`) | 0 % | 5,068,103,201 | 0.0601 / 0.00192 | 0.0285 / 0.00169 | FAIL |
+| asymmetric int8, `in_proj_qkv` + `out_proj` + `v_proj` fp16 | 21.7 % | 5,882,832,315 | 0.0175 / 0.00135 | 0.0277 / 0.00123 | FAIL |
+| asymmetric int8, `in_proj_qkv` + `out_proj` fp16 | 21.2 % | 5,863,831,482 | 0.0174 / 0.00138 | not run | — |
+
+The two asymmetric sets are the ones an fp32 torch bisect chose under a rule written before any result
+(at most 25 % of the linear weights fp16, the target worst row ≤ 0.006 met by none; the two best by mean
+taken). The one with `v_proj` fp16 passes the fixture and fails the held-out set on one `score` row, the row that also
+breaks int8lin and is fp16's held-out worst. A set built before the rule's last stage (layers 0–3 and
+`out_proj` fp16) fails both (0.0272, 0.0293). The quantizer's `qscheme` accepts `symmetric`,
+`asymmetric` and `symmetric_with_clipping`; block 16 and an int8 embedding table also compile
+([`gate-kev-4b-int8.json`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/gate-kev-4b-int8.json)).
+
+**What the compiled asset holds.** `--expect-frequent-reshapes` adds an fp16 copy of every linear weight:
+for the unrolled graph the h19p asset of the fp16 bundle is 15,557,759,365 bytes with it and 8,413,701,860
+without, and the int8 sets above compile to 13,026,670,589 and 13,007,670,689 bytes with it. This
+release's Mac h16c asset is 15,552,786,962 bytes against an 8,412,500,461-byte `main.mlirb`. Without the
+flag the runtime specializes the unrolled graph again for every new position length: 9.6–14.6 s per new
+length on the Mac GPU (calls at a length already seen: 46.4 ms median, 33 rows); an int8 asset without the
+flag had not finished specializing its opening length after 12 minutes. A Mac gate worker on the efr fp16
+asset of the unrolled graph showed up to 16.6 GB rss, up to 16.4 GB of it clean mapped file, and a
+phys_footprint of at most 1.15 GB.
+
+## ⬇️ Bundle
+
+[mlboydaisuke/Kev-4B-CoreAI](https://huggingface.co/mlboydaisuke/Kev-4B-CoreAI), one folder under
+`gpu-pipelined/`:
+
+| file | what | bytes | sha256 |
+|---|---|---:|---|
+| `kev_4b_decode_fp16_metal_pf128.aimodel/main.mlirb` | the decoder, fp16 | 8,412,500,461 | `da529dd4…413eb17f` |
+| `metadata.json` | `kind: decision-backbone`, the readout contract, the gate | 12,643 | `e641e44f…c9039b08` |
+| `tokenizer/tokenizer.json` | the base model's, verbatim | 12,807,196 | `fe000e3e…d50d2927` |
+| `tokenizer/tokenizer_config.json` | the base model's, verbatim | 16,713 | `3891e840…b6b9f89c` |
+| `head/head.safetensors` | the pointer head: q / k weight `[256, 2560]` and bias, fp32 | 5,245,304 | `36c392f7…54157106` |
+| `head/kev_head.json` | head size, scale, temperature, delimiter ids, provenance | 2,038 | `5ff071ff…cfba7416` |
+
+The repository root carries the base model's `config.json`, the adapter's `adapter_config.json` and the
+Apache-2.0 `LICENSE`, verbatim. `SHA256SUMS` lists every file. The Hub revision is recorded here after the
+upload.
+
+No AOT asset ships: the Swift runtime's specialization of the `.aimodel` equals the AOT asset bit for bit
+(above). To compile one for the Mac anyway (15.55 GB, the flag is required):
+
+```bash
+xcrun coreai-build compile kev_4b_decode_fp16_metal_pf128.aimodel --output aot --preferred-compute gpu \
+    --platform macOS --architecture h16c --expect-frequent-reshapes
+```
+
+## Use it
+
+Swift, with the [`Kev`](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/Kev) package (macOS 27; the system CoreAI framework, Accelerate and
+swift-transformers' tokenizer), on a download of the repository:
+
+```swift
+import Kev
+
+let bundle = URL(filePath: "Kev-4B-CoreAI/gpu-pipelined/kev_4b_decode_fp16_metal_pf128")
+let kev = try await KevDecider(bundle: bundle)              // the .aimodel, specialized here (GPU, frequent reshapes)
+let response = try await kev.decide(requestJSON: requestData, shared: true)   // shared: the state's whole calls run once
+
+// questions that arrive later, on the same state
+let prepared = try await kev.prepare(state: try JSONParser.parse(stateData))       // the state's whole calls, once
+let later = try await kev.decide(prepared: prepared, questionsJSON: questionsData)  // only the questions' rows run
+```
+
+Loading the bundle specializes the graph once (16.6 s on the M4 Max) and caches it; later loads read the cache.
+The Mac CLI: `kev run --bundle Kev-4B-CoreAI/gpu-pipelined/kev_4b_decode_fp16_metal_pf128 --asset jit
+--request req.json --shared --out resp.json` (`swift build -c release --package-path apps/Kev`). The
+request shape and the Python read-out (`conversion/kev/decide.py run --model kev-4b …`, AOT assets only)
+are as in the [Kev-0.8B card](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-0.8b/README.md#use-it).
+
+## Reproduce
+
+Environment: the zoo overlay venv (coreai-core 1.0.0b2, coreai-torch 0.4.1, torch 2.9.0, transformers
+4.57.6), Xcode 27.0 RC; the author's code (the oracle and the merge) in its own venv from the author's
+`uv.lock` at tag `kev-1.0`. Order and flags: [`conversion/kev/README.md`](https://github.com/john-rocky/coreai-model-zoo/blob/main/conversion/kev/README.md).
+
+```bash
+(cd $ZOO_WORK_ROOT/_kev/kev-src && python scripts/merge_lora_checkpoint.py \
+    --lora jaredpalmer/kev-4b@591dcb5bd6d05eb0b5131ea6608f93f10243335c --out $ZOO_WORK_ROOT/_kev/merged/kev-4b-v1.0)
+# the decoder (--aot adds the h16c .aimodelc the Python gates load)
+python conversion/kev/export_decoder.py fp16 --model kev-4b --gdn-scan metal --prefill-chunk 128 --aot
+# the published bundle: the gated .aimodel bytes and their metadata, without exporting again
+python conversion/kev/export_decoder.py fp16 --model kev-4b --gdn-scan metal --prefill-chunk 128 --metadata-only \
+    --from-aimodel <bundles>/kev_4b_decode_fp16_metal_pf128/kev_4b_decode_fp16_metal_pf128.aimodel \
+    --bundle-dir <ship>/kev_4b_decode_fp16_metal_pf128
+```
+
+Recipe: [`recipe.toml`](https://github.com/john-rocky/coreai-model-zoo/blob/main/models/kev-4b/recipe.toml). Port notes: [`knowledge/kev-port.md`](https://github.com/john-rocky/coreai-model-zoo/blob/main/knowledge/kev-port.md).
+
+## Other formats
+
+On the Hub (2026-10-04), not run here:
+
+- ONNX: [`midudev/kev-4b-ONNX`](https://huggingface.co/midudev/kev-4b-ONNX) (ONNX Runtime Web, int4
+  weights, the pointer head as `head.bin`);
+  [`onnx-community/kev-4b-ONNX`](https://huggingface.co/onnx-community/kev-4b-ONNX) (Transformers.js,
+  `q4` / `q4f16`, the pointer head in the graph; its card names Qwen/Qwen3-4B-Base as the base, the
+  Qwen3 generation of Kev-4B before Kev 1.0).
+- GGUF: [`ggml-org/Kev-4B-GGUF`](https://huggingface.co/ggml-org/Kev-4B-GGUF) ("a decision model, to be
+  used via `/v1/systemone` API", per its card).
+- MLX: [`RoderickQiu/kev-4b-mlx-8bit`](https://huggingface.co/RoderickQiu/kev-4b-mlx-8bit) (merged, 8-bit,
+  `head.pt` fp32) and [`aselea/Kev-4B-MLX-Serve-8bit`](https://huggingface.co/aselea/Kev-4B-MLX-Serve-8bit)
+  (for mlx-serve, the head as `kev_head.safetensors`); the author's repository also serves Kev through
+  MLX on Apple silicon.
+
+No other Core AI conversion of Kev was listed on the Hub on 2026-10-04.
+
+## License
+
+Apache-2.0: the adapter and the head (jaredpalmer/kev-4b) and the base (Qwen/Qwen3.5-4B-Base); the bundle
+inherits it, and the Hugging Face repository carries the base's `LICENSE`. The author's package runs only
+in the oracle and the merge, at gate time. The fixture file's terms are Kev-0.8B's: SemIf authored144
+under MIT with its notice, transfer-v4 records by reference, 11 of the 20 records written for the port
+published and 9 withheld (an invented name in each was found in use on the web).
+
+## Limits
+
+From the [author's card](https://huggingface.co/jaredpalmer/kev-4b): English only; text generation, chat
+and fully automated consequential decisions about people are out of scope; questions that need facts
+that are neither in the state nor general knowledge are out of scope; the temperature was fitted on the
+author's development rows (the card explains how to measure and refit it). This port adds:
+
+- A row holds at most 3,968 tokens, while the author's server accepts states up to 65,536.
+- The release is for the Mac.
+- A process's opening call is slower. On the Mac the opening 128-token call of each Python gate process
+  took 1,121–9,881 ms, against a median of 99.0 ms for its later calls.
+- If you export this decoder with a dynamic query length yourself, keep the call lengths to one or two
+  values. A process that cycles more lengths grows its memory until the `AIModel` is created again
+  (measured on Kev-0.8B on the Mac and the iPhone; which runtime layer keeps the memory is not isolated).

SHA256SUMS

@@ -0,0 +1,12 @@
+50cbab8a892c5f2993b8c7351a99182507472def3b1374558308605d99b86b32  LICENSE
+49f9ef81dab003ae1275678400962136811b0de05d2132fb9c8527a22d89c1fb  README.md
+8a05dfd6c5e7f61a62e6db5d8de8093a8dbbd105aa7ee7b5e7b73f6249f45617  adapter_config.json
+ddc63e1c717afa86c865bb5e01313d89d72bb53b97ad4a8a03ba8510c0621670  config.json
+36c392f75fa5eff2243b07664e1edb1dc4e771dfc9d145469ba175db54157106  gpu-pipelined/kev_4b_decode_fp16_metal_pf128/head/head.safetensors
+5ff071ffaa5b620999f44770a014b4b5b1da2a8225acbde10b412652cfba7416  gpu-pipelined/kev_4b_decode_fp16_metal_pf128/head/kev_head.json
+862ef6355b0169ce54a2b9a9159b9fcdbcdae2b227aa18075baf86fbcc9902e4  gpu-pipelined/kev_4b_decode_fp16_metal_pf128/kev_4b_decode_fp16_metal_pf128.aimodel/main.hash
+da529dd48f08a7535ca6b8dfb5640723641c2058daf7757a52a9a527413eb17f  gpu-pipelined/kev_4b_decode_fp16_metal_pf128/kev_4b_decode_fp16_metal_pf128.aimodel/main.mlirb
+2ab4242dec3222dd5338649a17fb1dc0cfc60f969f9d9bef27ce38fb84ec675b  gpu-pipelined/kev_4b_decode_fp16_metal_pf128/kev_4b_decode_fp16_metal_pf128.aimodel/metadata.json
+e641e44fc9d540614e7bc192e49df78c9d8a7d8e23309fe1bef499a5c9039b08  gpu-pipelined/kev_4b_decode_fp16_metal_pf128/metadata.json
+fe000e3ed39ed12b8d2481d527d44f93c65d37e87645d2dcc80d1bf9d50d2927  gpu-pipelined/kev_4b_decode_fp16_metal_pf128/tokenizer/tokenizer.json
+3891e840d7dc5fca0af33d3a25083a735e36fe06214e3f707024820cb6b9f89c  gpu-pipelined/kev_4b_decode_fp16_metal_pf128/tokenizer/tokenizer_config.json

adapter_config.json

@@ -0,0 +1,56 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3.5-4B-Base",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "kasa_config": null,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.21.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "in_proj_b",
+    "in_proj_qkv",
+    "gate_proj",
+    "v_proj",
+    "o_proj",
+    "k_proj",
+    "up_proj",
+    "in_proj_z",
+    "in_proj_a",
+    "out_proj",
+    "down_proj",
+    "q_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "FEATURE_EXTRACTION",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

config.json

@@ -0,0 +1,104 @@
+{
+    "architectures": [
+        "Qwen3_5ForConditionalGeneration"
+    ],
+    "image_token_id": 248056,
+    "model_type": "qwen3_5",
+    "text_config": {
+        "attention_bias": false,
+        "attention_dropout": 0.0,
+        "attn_output_gate": true,
+        "dtype": "bfloat16",
+        "eos_token_id": 248044,
+        "full_attention_interval": 4,
+        "head_dim": 256,
+        "hidden_act": "silu",
+        "hidden_size": 2560,
+        "initializer_range": 0.02,
+        "intermediate_size": 9216,
+        "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,
+        "max_position_embeddings": 262144,
+        "mlp_only_layers": [],
+        "model_type": "qwen3_5_text",
+        "mtp_num_hidden_layers": 1,
+        "mtp_use_dedicated_embeddings": false,
+        "num_attention_heads": 16,
+        "num_hidden_layers": 32,
+        "num_key_value_heads": 4,
+        "rms_norm_eps": 1e-06,
+        "tie_word_embeddings": true,
+        "use_cache": true,
+        "vocab_size": 248320,
+        "mamba_ssm_dtype": "float32",
+        "rope_parameters": {
+            "mrope_interleaved": true,
+            "mrope_section": [
+                11,
+                11,
+                10
+            ],
+            "rope_type": "default",
+            "rope_theta": 10000000,
+            "partial_rotary_factor": 0.25
+        }
+    },
+    "tie_word_embeddings": true,
+    "transformers_version": "4.57.0.dev0",
+    "video_token_id": 248057,
+    "vision_config": {
+        "deepstack_visual_indexes": [],
+        "depth": 24,
+        "hidden_act": "gelu_pytorch_tanh",
+        "hidden_size": 1024,
+        "in_channels": 3,
+        "initializer_range": 0.02,
+        "intermediate_size": 4096,
+        "model_type": "qwen3_5",
+        "num_heads": 16,
+        "num_position_embeddings": 2304,
+        "out_hidden_size": 2560,
+        "patch_size": 16,
+        "spatial_merge_size": 2,
+        "temporal_patch_size": 2
+    },
+    "vision_end_token_id": 248054,
+    "vision_start_token_id": 248053
+}

gpu-pipelined/kev_4b_decode_fp16_metal_pf128/head/head.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36c392f75fa5eff2243b07664e1edb1dc4e771dfc9d145469ba175db54157106
+size 5245304

gpu-pipelined/kev_4b_decode_fp16_metal_pf128/head/kev_head.json

@@ -0,0 +1,67 @@
+{
+ "model": "kev-4b",
+ "hidden_size": 2560,
+ "head_dim": 256,
+ "scale": 0.0625,
+ "temperature": 2.406050072164233,
+ "formula": "z_k = ((k.weight @ h_opt_k + k.bias) . (q.weight @ h_decide + q.bias)) * scale; p = softmax_k(z / temperature)",
+ "readout": {
+  "decide": "hidden state at the row's last token (<decide>)",
+  "option_k": "hidden state at option k's </opt> token",
+  "hidden": "the decoder's last_hidden_state (after the final RMSNorm), fp32"
+ },
+ "row": "one causal row per question: [state] + user_tokens(render(state)) + [q] + user_tokens(render(instructions)) + for each option ([opt] + user_tokens(option) + [opt_end]) + [decide]; positions 0..L-1; fresh state per row",
+ "user_tokens": "re.sub(r'<\\|([A-Za-z0-9_]+)\\|>', r'<¦\\1¦>', text), tokenized without special tokens",
+ "delimiters": {
+  "state": {
+   "token": "<|fim_prefix|>",
+   "id": 248060
+  },
+  "q": {
+   "token": "<|fim_middle|>",
+   "id": 248061
+  },
+  "opt": {
+   "token": "<|box_start|>",
+   "id": 248049
+  },
+  "opt_end": {
+   "token": "<|box_end|>",
+   "id": 248050
+  },
+  "decide": {
+   "token": "<|fim_suffix|>",
+   "id": 248062
+  }
+ },
+ "pad": {
+  "token": "<|endoftext|>",
+  "id": 248044
+ },
+ "base": {
+  "repo": "Qwen/Qwen3.5-4B-Base",
+  "revision": "1001bb4d826a52d1f399e183466143f4da7b741b"
+ },
+ "adapter": {
+  "repo": "jaredpalmer/kev-4b",
+  "tag": "v1.0",
+  "revision": "591dcb5bd6d05eb0b5131ea6608f93f10243335c",
+  "resolved_commit": "6cfce5c2fa4b4bd64026336ab649c5ca78857d52",
+  "adapter_sha256": "90e817356246e7f18bfa7ca3d31794cd4fbeb3332a66a84cb51d9ceae925f2b2",
+  "head_pt_sha256": "dd633435998ecc751ac538717a3742e32149500fabf7d7276287dbf0693f347c"
+ },
+ "head_safetensors_sha256": "36c392f75fa5eff2243b07664e1edb1dc4e771dfc9d145469ba175db54157106",
+ "temperature_fit": {
+  "rows": "runs/release/kev-4b-r10/development/rows.json",
+  "n": 1264,
+  "method": "min micro mean NLL over a 121-point log grid 0.25..4"
+ },
+ "record0_check": [
+  {
+   "record": "tv4_000",
+   "qid": "answer",
+   "bit_equal": true,
+   "max_abs_diff": 0.0
+  }
+ ]
+}

gpu-pipelined/kev_4b_decode_fp16_metal_pf128/kev_4b_decode_fp16_metal_pf128.aimodel/main.hash

@@ -0,0 +1 @@
+�R�ԏ�S\��ߵd#d X��uzR��'A>�

gpu-pipelined/kev_4b_decode_fp16_metal_pf128/kev_4b_decode_fp16_metal_pf128.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da529dd48f08a7535ca6b8dfb5640723641c2058daf7757a52a9a527413eb17f
+size 8412500461

gpu-pipelined/kev_4b_decode_fp16_metal_pf128/kev_4b_decode_fp16_metal_pf128.aimodel/metadata.json

@@ -0,0 +1,5 @@
+{
+  "creationDate" : "20261004T000304Z",
+  "producer" : "coreai-core 1.0.0b2",
+  "assetVersion" : "2.0"
+}

gpu-pipelined/kev_4b_decode_fp16_metal_pf128/metadata.json

@@ -0,0 +1,276 @@
+{
+  "metadata_version": "0.2",
+  "kind": "decision-backbone",
+  "name": "kev_4b_decode_fp16_metal_pf128",
+  "assets": {
+    "main": "kev_4b_decode_fp16_metal_pf128.aimodel"
+  },
+  "language": {
+    "tokenizer": "Qwen/Qwen3.5-4B-Base",
+    "vocab_size": 248320,
+    "max_context_length": 4096,
+    "embedded_tokenizer": true,
+    "prefill_chunk": 128,
+    "static_inputs": [],
+    "output": "hidden [1, 128, 2560] fp16, every position",
+    "function_map": {
+      "main": [
+        "main"
+      ]
+    },
+    "tokenizer_revision": "1001bb4d826a52d1f399e183466143f4da7b741b"
+  },
+  "source": {
+    "model_definition": "torch (coreai_models.models.macos.qwen3_5, conversion/kev/qwen3_5_kev_decoder.py)",
+    "adapter": {
+      "repo": "jaredpalmer/kev-4b",
+      "tag": "v1.0",
+      "tag_target": "591dcb5bd6d05eb0b5131ea6608f93f10243335c",
+      "resolved_commit": "6cfce5c2fa4b4bd64026336ab649c5ca78857d52",
+      "weights_identical_to": "139fdd94f1b6a6ad80cc15e08fcb99cac885a101 (the card's weight revision; same LFS oids)",
+      "adapter_model_safetensors_sha256": "90e817356246e7f18bfa7ca3d31794cd4fbeb3332a66a84cb51d9ceae925f2b2",
+      "head_pt_sha256": "dd633435998ecc751ac538717a3742e32149500fabf7d7276287dbf0693f347c",
+      "license": "apache-2.0"
+    },
+    "base": {
+      "repo": "Qwen/Qwen3.5-4B-Base",
+      "revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
+      "safetensors_sha256": {
+        "model.safetensors-00001-of-00002.safetensors": "df547074dce70532a0493e5433152bd17a65efb89088cfabc2e7e2371a93d712",
+        "model.safetensors-00002-of-00002.safetensors": "590fbaac095dd31db886c322d9d2f7df47777966391acf306ddddc3e4e3a15ef"
+      },
+      "license": "apache-2.0"
+    },
+    "merge": {
+      "tool": "scripts/merge_lora_checkpoint.py (https://github.com/jaredpalmer/kev tag kev-1.0, commit 6b719c3c3f367295f6ef336f4f751cf5ff970abc)",
+      "formula": "W + (B @ A) * alpha / r (alpha / r = 2.0) on every adapted weight, in fp32, written fp32 (12 target kinds, 248 adapted tensors); every other weight unchanged",
+      "weights_sha256": "904380cbf0be134e122f2abd7c7bab52f1e2a0f373d3a2e7947a3025e882f3f3",
+      "model_safetensors_sha256": {
+        "model-00001-of-00004.safetensors": "45a919c6e0c4a907324e8735cd3be10ab8bdad2ba67fa5844367e0fe3a4bd18f",
+        "model-00002-of-00004.safetensors": "9e35c50caf92e662b03a092fc6b57db69683f9ae4076661a2aef962ea175aaab",
+        "model-00003-of-00004.safetensors": "c9036c7fed9c87d419f81435955b4a84e6803c1d8b7029e37cd9979aaf6815f3",
+        "model-00004-of-00004.safetensors": "47cde07c0ef1909177ea7c6f6bc75033443ca186f8567d72a9ef5bfd000d2b89"
+      },
+      "tensors": 426
+    },
+    "hf_model_id": "jaredpalmer/kev-4b",
+    "hf_revision": "6cfce5c2fa4b4bd64026336ab649c5ca78857d52",
+    "tokenizer": {
+      "repo": "Qwen/Qwen3.5-4B-Base",
+      "revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
+      "files_sha256": {
+        "tokenizer.json": "fe000e3ed39ed12b8d2481d527d44f93c65d37e87645d2dcc80d1bf9d50d2927",
+        "tokenizer_config.json": "3891e840d7dc5fca0af33d3a25083a735e36fe06214e3f707024820cb6b9f89c"
+      },
+      "shipped_in": "tokenizer/ (verbatim copies of the two files; embedded_tokenizer)"
+    },
+    "hf_revision_note": "the commit the Hub resolves the tag v1.0 to; the adapter and head.pt are the LFS objects named in source.adapter. The exporter read the merged checkpoint (source.merge) through a local cache id, which is not recorded here."
+  },
+  "compression": null,
+  "decision": {
+    "output": "hidden [1, S, 2560] per call: the final-norm hidden state at every position (no vocabulary head in the graph)",
+    "readout": "one row per question, fresh zero states per row; a row of T ids runs as ceil(T / 128) calls of 'main' (static S = 128): call k gets ids[128k : 128k + 128] with position_ids 0..128k+127; 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, 128, 2560] of every call, concatenated and cut to T, are the backbone's final-norm last_hidden_state [T, 2560] the head reads. A row fits when ceil(T / 128) * 128 <= 4095 (the position dim's upper bound; the KV sequence dim is allocated at 4096).",
+    "row": {
+      "source": "the author's kev.model.encode + rows_of (row form, https://github.com/jaredpalmer/kev tag kev-1.0); conversion/kev/oracle_kev.py records the oracle's rows",
+      "layout": "[<state>] + user_tokens(render(state)) + [<q>] + user_tokens(render(instructions)) + for each option ([<opt>] + user_tokens(option_text) + [</opt>]) + [<decide>]",
+      "delimiters": {
+        "state": {
+          "token": "<|fim_prefix|>",
+          "id": 248060
+        },
+        "q": {
+          "token": "<|fim_middle|>",
+          "id": 248061
+        },
+        "opt": {
+          "token": "<|box_start|>",
+          "id": 248049
+        },
+        "opt_end": {
+          "token": "<|box_end|>",
+          "id": 248050
+        },
+        "decide": {
+          "token": "<|fim_suffix|>",
+          "id": 248062
+        }
+      },
+      "pad": {
+        "token": "<|endoftext|>",
+        "id": 248044
+      },
+      "bos": null,
+      "positions": "0..L-1 (one causal row = the state ids then the question's branch ids)",
+      "user_tokens": "re.sub(r'<\\|([A-Za-z0-9_]+)\\|>', r'<¦\\1¦>', text), then tokenized without special tokens (user text can never produce a delimiter)",
+      "tokenizer": "tokenizer/ (the base model's pinned tokenizer.json + tokenizer_config.json)",
+      "readout_positions": {
+        "decide": "the row's last token (<decide>)",
+        "options": "each option's </opt> token, in option order"
+      },
+      "limits_author_serving": "state <= 65,536 tokens (the <state> token included), a row <= 73,728; this graph: one row <= the readout bound above"
+    },
+    "request": {
+      "source": "kev.api.to_record (SystemOne request -> internal record)",
+      "render": "None -> ''; str / int / float / bool -> str(v); array -> one line per item '- ' + render(item) (nested items indented by two spaces per level); object -> one line per key 'key: value', or 'key:' then the nested render indented by two spaces when the value is an object or an array",
+      "state": "render(state)",
+      "instructions": "render(instructions) (omitted -> '': the <q> token is followed by the options)",
+      "option_text": "name if description is None or '' else 'name: ' + render(description)",
+      "options": {
+        "noul": [
+          "option_text('no', criteria.false)",
+          "option_text('yes', criteria.true)"
+        ],
+        "choice": "option_text(name, description) for each criteria entry, in criteria order",
+        "score": "render(level) for each level, in order"
+      },
+      "question_keys": {
+        "noul": [
+          "false",
+          "true"
+        ],
+        "choice": "the criteria names, in order",
+        "score": "'0'..'n-1'"
+      },
+      "max_options": 255
+    },
+    "head": {
+      "files": [
+        "head/head.safetensors",
+        "head/kev_head.json"
+      ],
+      "formula": "z_k = ((k.weight @ h_opt_k + k.bias) . (q.weight @ h_decide + q.bias)) * 0.0625; p = softmax_k(z / temperature)",
+      "scale": 0.0625,
+      "temperature": "kev_head.json temperature (2.406050072164233, the author's calibration in head.pt)",
+      "code": "kev.model.PointerHead.forward (q, k = nn.Linear(2560, 256)); weights from head.pt, fp32",
+      "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, after the temperature",
+    "response": {
+      "shape": "SystemOne-compatible: {model, answers: {question_id: answer}, usage: {input_tokens, output_tokens}, latency_ms} (kev.serve.Server._body)",
+      "noul": "{type: 'noul', noul: p(true)}",
+      "choice": "{type: 'choice', choice: the key of the first max p, confidence: (p_max - 1/K) / (1 - 1/K) (1 when K = 1), probabilities: {key: p}}",
+      "score": "{type: 'score', score: sum(i * p_i), legend: {key: render(level)}, probabilities: {key: p}, confidence: max(0, 1 - E|level - mode| / D), D = mean |i - (L-1)/2| over the L levels, mode = the first most likely level (1 when L = 1)}",
+      "normalize": "the confidences normalize p to sum 1 first (all zeros -> uniform)",
+      "rounding": "4 decimals (kev.api.round_prob)",
+      "usage": "input_tokens = the encoded request's tokens (the author's packed form: state once + every branch); output_tokens = tokens of json.dumps(answers) (kev.api.output_tokens)"
+    }
+  },
+  "gdn_scan": {
+    "form": "metal",
+    "kernel": "qwen3_5_gdn_chunk_s128",
+    "chunk_max": 128,
+    "kernel_source": "coreai_models.models.macos.qwen3_5_gdn_metal.build_gdn_chunk_kernel (overlay, unchanged)",
+    "threads_per_grid": [
+      128,
+      32,
+      1
+    ],
+    "threads_per_thread_group": [
+      128,
+      1,
+      1
+    ]
+  },
+  "compilation": {
+    "date": "2026-10-04T05:42:03.402948+00:00",
+    "targets": []
+  },
+  "gate": {
+    "bar": {
+      "argmax": "every question whose oracle top-2 margin is above 0.02 (near-ties listed apart)",
+      "max_abs_dp": 0.02,
+      "mean_of_run_mean_abs_dp": 0.002,
+      "reset": "bit-equal in every process",
+      "finite": true
+    },
+    "oracle": "the author's fp32 checkpoint on the CPU (kev.model.admit + the row-form forward, threads 1)",
+    "graph": "the AOT h16c .aimodelc of this bundle's .aimodel, Python runtime, SpecializationOptions.default(), 128 ids per call",
+    "fixture_434": {
+      "result": "PASS",
+      "rows": 434,
+      "argmax_non_near_tie": "422/422",
+      "argmax_near_tie": "10/12",
+      "max_abs_dp": 0.015306,
+      "mean_of_run_mean_abs_dp": 0.0007686,
+      "worst_row": {
+        "id": "semif_1105577da4c8dad4609d",
+        "q": 0,
+        "max_abs_dp": 0.015306353569030762
+      },
+      "reset_bit_equal_all_processes": true,
+      "finite": true,
+      "oracle_sha256": "eb3031df83a75b746beec3480a4bbb47787e13f8a12412ff135feb9f609c70f9",
+      "when": "2026-10-04T02:09:21+00:00",
+      "lane_file_sha256": "41dca9cbed8d58bf331480137dee62f212449769c2e723b2f7d79ca7f43ab6b4",
+      "transcript": "coreai-model-zoo models/kev-4b/gate-kev-4b-readout.json"
+    },
+    "heldout_130": {
+      "result": "PASS",
+      "rows": 130,
+      "argmax_non_near_tie": "126/126",
+      "argmax_near_tie": "3/4",
+      "max_abs_dp": 0.00961,
+      "mean_of_run_mean_abs_dp": 0.0006898,
+      "worst_row": {
+        "id": "tv4sh_11",
+        "q": 0,
+        "max_abs_dp": 0.009610295295715332
+      },
+      "reset_bit_equal_all_processes": true,
+      "finite": true,
+      "oracle_sha256": "dfa101a559450a8b75c140bafd120728561273908a527ad89ce44d5558d31c59",
+      "when": "2026-10-04T02:10:22+00:00",
+      "lane_file_sha256": "96f097809ee73d5898cdd64614c05033c253a4012175e15a1d818046585d1738",
+      "transcript": "coreai-model-zoo models/kev-4b/gate-kev-4b-readout.json"
+    },
+    "red_arms_v2": {
+      "result": "RED (every arm)",
+      "transcript": "coreai-model-zoo models/kev-4b/gate-kev-4b-readout.json"
+    },
+    "swift": {
+      "result": "PASS",
+      "hidden_sha256_equal_python_gate": {
+        "fixture": 434,
+        "heldout": 130
+      },
+      "shared_hidden_bit_equal_direct": {
+        "fixture": 434,
+        "heldout": 130
+      },
+      "jit_vs_aot": {
+        "records": 10,
+        "rows": 25,
+        "hidden_sha256_equal": 25,
+        "p_bit_equal": 25
+      },
+      "transcript": "coreai-model-zoo models/kev-4b/gate-kev-4b-swift.json"
+    },
+    "host": {
+      "transcript": "coreai-model-zoo models/kev-4b/gate-kev-4b-host.json"
+    },
+    "fixture": "coreai-model-zoo models/kev-4b/fixtures-kev-4b.json",
+    "devices": "Mac only in this release"
+  },
+  "metadata_of_record": {
+    "exported_metadata_json_sha256": "dcdd0ea1c4ea4369282176db2f937779aab00b944b4f26034e9314b192d88e87",
+    "edits": [
+      "language.tokenizer = the base repo id (was the local id of the merged checkpoint) + tokenizer_revision",
+      "source.hf_model_id / hf_revision = the Kev checkpoint (was the local HF-cache id of the merged weights)",
+      "source.tokenizer.shipped_in added",
+      "gate: the published form's gate numbers and the zoo transcripts that hold them"
+    ],
+    "host_read_keys_unchanged": [
+      "name",
+      "kind",
+      "assets",
+      "language.vocab_size",
+      "language.max_context_length",
+      "language.prefill_chunk",
+      "language.query_len_range",
+      "language.query_len_call_max",
+      "language.query_len_multiple",
+      "decision"
+    ],
+    "written": "2026-10-04 (round 17)"
+  }
+}

gpu-pipelined/kev_4b_decode_fp16_metal_pf128/tokenizer/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe000e3ed39ed12b8d2481d527d44f93c65d37e87645d2dcc80d1bf9d50d2927
+size 12807196

gpu-pipelined/kev_4b_decode_fp16_metal_pf128/tokenizer/tokenizer_config.json

@@ -0,0 +1,305 @@
+{
+    "add_prefix_space": false,
+    "added_tokens_decoder": {
+        "248044": {
+            "content": "<|endoftext|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248045": {
+            "content": "<|im_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248046": {
+            "content": "<|im_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248047": {
+            "content": "<|object_ref_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248048": {
+            "content": "<|object_ref_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248049": {
+            "content": "<|box_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248050": {
+            "content": "<|box_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248051": {
+            "content": "<|quad_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248052": {
+            "content": "<|quad_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248053": {
+            "content": "<|vision_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248054": {
+            "content": "<|vision_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248055": {
+            "content": "<|vision_pad|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248056": {
+            "content": "<|image_pad|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248057": {
+            "content": "<|video_pad|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248058": {
+            "content": "<tool_call>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248059": {
+            "content": "</tool_call>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248060": {
+            "content": "<|fim_prefix|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248061": {
+            "content": "<|fim_middle|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248062": {
+            "content": "<|fim_suffix|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248063": {
+            "content": "<|fim_pad|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248064": {
+            "content": "<|repo_name|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248065": {
+            "content": "<|file_sep|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248066": {
+            "content": "<tool_response>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248067": {
+            "content": "</tool_response>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248068": {
+            "content": "<think>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248069": {
+            "content": "</think>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": false
+        },
+        "248070": {
+            "content": "<|audio_start|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248071": {
+            "content": "<|audio_end|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248072": {
+            "content": "<tts_pad>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248073": {
+            "content": "<tts_text_bos>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248074": {
+            "content": "<tts_text_eod>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248075": {
+            "content": "<tts_text_bos_single>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        },
+        "248076": {
+            "content": "<|audio_pad|>",
+            "lstrip": false,
+            "normalized": false,
+            "rstrip": false,
+            "single_word": false,
+            "special": true
+        }
+    },
+    "additional_special_tokens": [
+        "<|im_start|>",
+        "<|im_end|>",
+        "<|object_ref_start|>",
+        "<|object_ref_end|>",
+        "<|box_start|>",
+        "<|box_end|>",
+        "<|quad_start|>",
+        "<|quad_end|>",
+        "<|vision_start|>",
+        "<|vision_end|>",
+        "<|vision_pad|>",
+        "<|image_pad|>",
+        "<|video_pad|>"
+    ],
+    "bos_token": null,
+    "chat_template": "{%- set image_count = namespace(value=0) %}\n{%- set video_count = namespace(value=0) %}\n{%- macro render_content(content, do_vision_count, is_system_content=false) %}\n    {%- if content is string %}\n        {{- content }}\n    {%- elif content is iterable and content is not mapping %}\n        {%- for item in content %}\n            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}\n                {%- if is_system_content %}\n                    {{- raise_exception('System message cannot contain images.') }}\n                {%- endif %}\n                {%- if do_vision_count %}\n                    {%- set image_count.value = image_count.value + 1 %}\n                {%- endif %}\n                {%- if add_vision_id %}\n                    {{- 'Picture ' ~ image_count.value ~ ': ' }}\n                {%- endif %}\n                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}\n            {%- elif 'video' in item or item.type == 'video' %}\n                {%- if is_system_content %}\n                    {{- raise_exception('System message cannot contain videos.') }}\n                {%- endif %}\n                {%- if do_vision_count %}\n                    {%- set video_count.value = video_count.value + 1 %}\n                {%- endif %}\n                {%- if add_vision_id %}\n                    {{- 'Video ' ~ video_count.value ~ ': ' }}\n                {%- endif %}\n                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}\n            {%- elif 'text' in item %}\n                {{- item.text }}\n            {%- else %}\n                {{- raise_exception('Unexpected item type in content.') }}\n            {%- endif %}\n        {%- endfor %}\n    {%- elif content is none or content is undefined %}\n        {{- '' }}\n    {%- else %}\n        {{- raise_exception('Unexpected content type.') }}\n    {%- endif %}\n{%- endmacro %}\n{%- if not messages %}\n    {{- raise_exception('No messages provided.') }}\n{%- endif %}\n{%- if tools and tools is iterable and tools is not mapping %}\n    {{- '<|im_start|>system\\n' }}\n    {{- \"# Tools\\n\\nYou have access to the following functions:\\n\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\" }}\n    {{- '\\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>' }}\n    {%- if messages[0].role == 'system' %}\n        {%- set content = render_content(messages[0].content, false, true)|trim %}\n        {%- if content %}\n            {{- '\\n\\n' + content }}\n        {%- endif %}\n    {%- endif %}\n    {{- '<|im_end|>\\n' }}\n{%- else %}\n    {%- if messages[0].role == 'system' %}\n        {%- set content = render_content(messages[0].content, false, true)|trim %}\n        {{- '<|im_start|>system\\n' + content + '<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" %}\n        {%- set content = render_content(message.content, false)|trim %}\n        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}\n            {%- set ns.multi_step_tool = false %}\n            {%- set ns.last_query_index = index %}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if ns.multi_step_tool %}\n    {{- raise_exception('No user query found in messages.') }}\n{%- endif %}\n{%- for message in messages %}\n    {%- set content = render_content(message.content, true)|trim %}\n    {%- if message.role == \"system\" %}\n        {%- if not loop.first %}\n            {{- raise_exception('System message must be at the beginning.') }}\n        {%- endif %}\n    {%- elif message.role == \"user\" %}\n        {{- '<|im_start|>' + message.role + '\\n' + content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set reasoning_content = '' %}\n        {%- if message.reasoning_content is string %}\n            {%- set reasoning_content = message.reasoning_content %}\n        {%- else %}\n            {%- if '</think>' in content %}\n                {%- set reasoning_content = content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n                {%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n            {%- endif %}\n        {%- endif %}\n        {%- set reasoning_content = reasoning_content|trim %}\n        {%- if loop.index0 > ns.last_query_index %}\n            {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content + '\\n</think>\\n\\n' + content }}\n        {%- else %}\n            {{- '<|im_start|>' + message.role + '\\n' + content }}\n        {%- endif %}\n        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if tool_call.function is defined %}\n                    {%- set tool_call = tool_call.function %}\n                {%- endif %}\n                {%- if loop.first %}\n                    {%- if content|trim %}\n                        {{- '\\n\\n<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n                    {%- else %}\n                        {{- '<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n                    {%- endif %}\n                {%- else %}\n                    {{- '\\n<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n                {%- endif %}\n                {%- if tool_call.arguments is defined %}\n                    {%- for args_name, args_value in tool_call.arguments|items %}\n                        {{- '<parameter=' + args_name + '>\\n' }}\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 %}\n                        {{- args_value }}\n                        {{- '\\n</parameter>\\n' }}\n                    {%- endfor %}\n                {%- endif %}\n                {{- '</function>\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.previtem and loop.previtem.role != \"tool\" %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- content }}\n        {{- '\\n</tool_response>' }}\n        {%- if not loop.last and loop.nextitem.role != \"tool\" %}\n            {{- '<|im_end|>\\n' }}\n        {%- elif loop.last %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- else %}\n        {{- raise_exception('Unexpected message role.') }}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n    {%- if enable_thinking is defined and enable_thinking is false %}\n        {{- '<think>\\n\\n</think>\\n\\n' }}\n    {%- else %}\n        {{- '<think>\\n' }}\n    {%- endif %}\n{%- endif %}",
+    "clean_up_tokenization_spaces": false,
+    "eos_token": "<|endoftext|>",
+    "errors": "replace",
+    "model_max_length": 262144,
+    "pad_token": "<|endoftext|>",
+    "split_special_tokens": false,
+    "tokenizer_class": "Qwen2Tokenizer",
+    "unk_token": null,
+    "add_bos_token": false,
+    "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+",
+    "extra_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|>"
+    }
+}