From ebdf740194d7cca40569b378daab36114b7012c0 Mon Sep 17 00:00:00 2001
From: Seto Elkahfi <1797197+setoelkahfi@users.noreply.github.com>
Date: Fri, 30 Oct 2026 11:45:30 +0100
Subject: [PATCH 1/3] Move onde-mistralrs to 0.9.5

Point every target at the published onde-mistralrs{,-core} 0.9.5 crates
instead of 0.9.3. Cargo.lock now resolves them from crates.io, with a
single onde-candle-core 0.11.0 and mio 1.2.3.

Upstream #2493 made Metal paged attention the default. Its KV block pool
does not fit in memory on iOS, tvOS, visionOS or watchOS, so both load
paths in engine.rs and the validation load in hf_cache.rs now call
without_paged_attn() on those targets. On tvOS, the auto device map
override is replaced by with_max_model_len(512).

Update the forked-dependencies and sdk-regressions skills with the
publish order and the patch-by-package-name trick for testing unpublished
fork changes.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---
 .agents/skills/forked-dependencies/SKILL.md |  11 +-
 .agents/skills/sdk-regressions/SKILL.md     |  14 ++-
 Cargo.lock                                  | 108 +++++++++++---------
 Cargo.toml                                  |  40 ++++----
 src/hf_cache.rs                             |  12 +++
 src/inference/engine.rs                     |  40 +++++---
 6 files changed, 143 insertions(+), 82 deletions(-)

diff --git a/.agents/skills/forked-dependencies/SKILL.md b/.agents/skills/forked-dependencies/SKILL.md
index 761c9e7..1531bd0 100644
--- a/.agents/skills/forked-dependencies/SKILL.md
+++ b/.agents/skills/forked-dependencies/SKILL.md
@@ -69,8 +69,8 @@ tqdm and mio entries.
 
 | Crate | Fork / source | Ref | Purpose | Consumed where |
 |-------|---------------|-----|---------|----------------|
-| mistral.rs (12 crates) | `setoelkahfi/mistral.rs` | branch `fix/all-platform-fixes` | All-platform fixes; **published** as `onde-mistralrs-*` | onde deps (registry) |
-| candle-core, candle-nn | `setoelkahfi/candle` | branch `onde/candle-0.11.0-27f20fea` (rev `27f20fea`) | Post-0.11.0 candle APIs; **published** as `onde-candle-{core,nn}` | mistral.rs fork deps (registry) |
+| mistral.rs (12 crates) | `setoelkahfi/mistral.rs` | branch `fix/all-platform-fixes` | All-platform fixes; **published** as `onde-mistralrs-*` (0.9.3 on crates.io; 0.9.5 synced with upstream, not yet published) | onde deps (git until 0.9.5 is published, then registry) |
+| candle-core, candle-nn | `setoelkahfi/candle` | branch `onde/candle-0.11.0-27f20fea` (rev `27f20fea`) | Post-0.11.0 candle APIs; **published** as `onde-candle-{core,nn}`. Upstream mistral.rs is now on candle rev `66a8cf18`, but the fork still compiles against this snapshot (host, iOS); `reduce.metal` changed in between, so watch Metal runtime behavior | mistral.rs fork deps (registry) |
 | tqdm | *none — upstream* | crates.io `tqdm 0.8.1` | Crossterm 0.29 (→ mio 1.x) is upstream as of 0.8.1 | mistral.rs fork dep (registry) |
 | sysctl | `setoelkahfi/sysctl-rs` | branch `feature/watchos` | watchOS (`target_os = "watchos"`) support | `[patch.crates-io]` in onde **and** mistral.rs fork |
 | mio | `setoelkahfi/mio` | branch `feature/visionos` | visionOS support for the `mio 0.8.11` line | **RETIRED** — patch removed from onde once it consumed onde-mistralrs 0.9.3 (crossterm 0.29 → mio 1.x). Fork kept for history/upstream PR. |
@@ -98,6 +98,13 @@ tqdm and mio entries.
   `-flash-attn`, `-metal-compile`, `-code-exec`, `-sandbox`, `-macros`,
   `-vision`, `-audio`, `-mcp`. The rename is a `package = "onde-..."` override on
   each workspace dep entry, so Rust source still `use`s `mistralrs_*` unchanged.
+- **Paged attention:** upstream #2493 made Metal paged attention the default.
+  onde turns it off on iOS, tvOS, visionOS and watchOS with
+  `.without_paged_attn()` (GGUF and UQFF builders in `engine.rs`, and the
+  validation load in `hf_cache.rs`); macOS keeps the default. Android is CPU
+  and unchanged.
+- **0.9.5 API change:** `with_device_mapping` now takes `DeviceMapMetadata`,
+  so the tvOS memory cap uses `.with_max_model_len(512)` instead.
 - **Publishing + bump flow:** see [[sdk-regressions]] (the 12-crate publish
   order, the candle Step 0, and the tqdm/core2 publish-verification snags).
 
diff --git a/.agents/skills/sdk-regressions/SKILL.md b/.agents/skills/sdk-regressions/SKILL.md
index 9ee49fc..b67a11a 100644
--- a/.agents/skills/sdk-regressions/SKILL.md
+++ b/.agents/skills/sdk-regressions/SKILL.md
@@ -422,7 +422,7 @@ target first.
 ### Step 0 — check if the candle snapshot needs an `onde-candle-*` publish
 
 Since 0.9.x, upstream mistral.rs pins `candle` to a git rev past the latest
-released candle (e.g. rev `27f20fea`, 19 commits past `0.11.0`, adding
+released candle (e.g. rev `27f20fea`; upstream has since moved to `66a8cf18`, which 0.9.5 does not need to compile, 19 commits past `0.11.0`, adding
 `QTensor::indexed_gemv`, `gemv_fused_shared_lhs`, `BarrierPool::execute_chunked`).
 `cargo publish` strips git pins, so publish verification resolves candle from
 crates.io and fails with E0599 "method not found" if the pinned rev has APIs the
@@ -456,6 +456,18 @@ candle-metal-kernels = { version = "0.11.0" }   # unchanged at rev -> official
 
 ### Step 1 — check if `onde-mistralrs` needs a new publish
 
+> Testing before publish: keep onde on the registry form
+> (`version = "0.9.5", package = "onde-mistralrs{,-core}"`) and patch by the real
+> package name from the command line, e.g. `cargo check --config
+> 'patch.crates-io.onde-mistralrs.git="https://github.com/setoelkahfi/mistral.rs"'
+> --config 'patch.crates-io.onde-mistralrs.branch="fix/all-platform-fixes"'`
+> (repeat for `onde-mistralrs-core`). A `[patch.crates-io]` rename of `mistralrs`
+> is silently ignored. `cargo package -p <the 12 crates>` in the fork verifies
+> the packaged crates against an overlay of their unpublished siblings. Don't
+> commit the Cargo.lock these produce. Since 0.9.5, Metal
+> paged attention is upstream's default (#2493), so onde calls
+> `.without_paged_attn()` on iOS/tvOS/visionOS/watchOS.
+
 First, see how far the fork has drifted from upstream:
 
 ```bash
diff --git a/Cargo.lock b/Cargo.lock
index eb13a5d..2e86816 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1318,17 +1318,6 @@ dependencies = [
  "serde_core",
 ]
 
-[[package]]
-name = "derive-new"
-version = "0.7.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2cdc8d50f426189eef89dac62fabfa0abb27d5cc008f25bf4156a0203325becc"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn 2.0.119",
-]
-
 [[package]]
 name = "derive_builder"
 version = "0.20.2"
@@ -3318,6 +3307,16 @@ dependencies = [
  "paste",
 ]
 
+[[package]]
+name = "metrics"
+version = "0.24.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "89550ee9f79e88fef3119de263694973a8adb26c21d75322164fb8c493039fe2"
+dependencies = [
+ "portable-atomic",
+ "rapidhash",
+]
+
 [[package]]
 name = "mime"
 version = "0.3.17"
@@ -3340,6 +3339,7 @@ version = "2.24.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "86886cf6dbf4e614b19c9a1eec9775f021869d7eadde0fc73921a81b90c9b4c9"
 dependencies = [
+ "indexmap 2.14.2",
  "memo-map",
  "serde",
  "serde_json",
@@ -3799,9 +3799,9 @@ dependencies = [
 
 [[package]]
 name = "onde-mistralrs"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "689f038ebc5d6d11e4880a24e50d62f91bbf21358b92c9b5cf7bf5e65af232bd"
+checksum = "0d9f41269feaffb3e452022f04a9cfb95859ed950889b6c1d010391531847260"
 dependencies = [
  "anyhow",
  "clap 4.6.6",
@@ -3827,9 +3827,9 @@ dependencies = [
 
 [[package]]
 name = "onde-mistralrs-audio"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "71b9a72504697c5b91e1d0448d67883e8d3046ad060200610c0484aef5dfaee5"
+checksum = "787f40176eda80bdef527227ba43a7d4dac3287a3d7e28af5b4072b92abee042"
 dependencies = [
  "anyhow",
  "apodize",
@@ -3839,9 +3839,9 @@ dependencies = [
 
 [[package]]
 name = "onde-mistralrs-code-exec"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4016265a19fa51ad973e99a6746904e432d3140b5c2ca14a41077edc6b8648c2"
+checksum = "b3851924d7b5dfceaa8309ada9d04d983a4b6cf8761117145e1a035011e4078f"
 dependencies = [
  "anyhow",
  "base64 0.22.1",
@@ -3859,9 +3859,9 @@ dependencies = [
 
 [[package]]
 name = "onde-mistralrs-core"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3e9266eb62ceb5a9cf8fb5342e6a9b78e286f47b670762559ee800a5a2db5132"
+checksum = "95d792d7cb5a4b78f4a448aa246bb6fbb8ea84a7958ce256173c12a83adbafb1"
 dependencies = [
  "ahash",
  "akin",
@@ -3878,9 +3878,9 @@ dependencies = [
  "chrono",
  "clap 4.6.6",
  "csv",
- "derive-new",
  "derive_more",
  "dirs",
+ "dispatch2",
  "either",
  "float8",
  "futures",
@@ -3899,6 +3899,7 @@ dependencies = [
  "libc",
  "llguidance",
  "lrtable",
+ "metrics",
  "mime_guess",
  "minijinja",
  "minijinja-contrib",
@@ -3910,11 +3911,11 @@ dependencies = [
  "onde-mistralrs-audio",
  "onde-mistralrs-code-exec",
  "onde-mistralrs-mcp",
+ "onde-mistralrs-metal-compile",
  "onde-mistralrs-paged-attn",
  "onde-mistralrs-quant",
  "onde-mistralrs-sandbox",
  "onde-mistralrs-vision",
- "onde-tqdm",
  "openai-harmony",
  "ordered-float",
  "parking_lot",
@@ -3938,10 +3939,12 @@ dependencies = [
  "serde-saphyr",
  "serde_json",
  "serde_plain",
+ "sha2",
  "statrs",
  "strum 0.27.2",
  "symphonia",
  "sysinfo",
+ "tempfile",
  "thiserror 2.0.20",
  "tokenizers 0.21.4",
  "tokio",
@@ -3950,6 +3953,7 @@ dependencies = [
  "toktrie",
  "toktrie_hf_tokenizers",
  "toml",
+ "tqdm",
  "tracing",
  "tracing-subscriber",
  "urlencoding",
@@ -3960,9 +3964,9 @@ dependencies = [
 
 [[package]]
 name = "onde-mistralrs-macros"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6e624ab1f0896a07ff437f4c49050c891918c0f01a008113c788b64e5f76c1d9"
+checksum = "8dd60c4cb0dc37105b335451643d075af0d293b8203df9bb038e88d83df8e475"
 dependencies = [
  "darling 0.23.0",
  "proc-macro2",
@@ -3972,9 +3976,9 @@ dependencies = [
 
 [[package]]
 name = "onde-mistralrs-mcp"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c8bcb140ed84c4dd2683f1533fc58905f959e7394843bc02a315bfdd3ffcbc25"
+checksum = "f5b5bf129c896c975a98575b9752291bb47b38a5bd3170944a8c171cc491af1b"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -3994,9 +3998,9 @@ dependencies = [
 
 [[package]]
 name = "onde-mistralrs-metal-compile"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "44109415be2971d311b403095d3c278443210d0ccc1c383a84ad44da5f5b1abf"
+checksum = "02e15b1154394bc9750f69c52b55de0295e6f7d3a3f3151f9ad8d88e066629af"
 dependencies = [
  "candle-metal-kernels",
  "objc2-metal",
@@ -4004,9 +4008,9 @@ dependencies = [
 
 [[package]]
 name = "onde-mistralrs-paged-attn"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6f93c8c04d15ebce014e80675387d269bcead85be063bf9afce8b353d64e61b8"
+checksum = "9ba6ae15cc8ce43c949fd038a8df0858e41dca38e4b4b77925aa04e6dc4aee84"
 dependencies = [
  "anyhow",
  "candle-metal-kernels",
@@ -4022,17 +4026,20 @@ dependencies = [
 
 [[package]]
 name = "onde-mistralrs-quant"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1d4ed48af14e3e5ff318a6ea6cefbc1e50ee753f143bd1ae6c0f188c33033d5a"
+checksum = "a25f0259f21cf04d86f18b3b0f48b072bcc4a989cadae513ca142ab1ccb0c60a"
 dependencies = [
+ "bytemuck",
  "byteorder",
  "candle-metal-kernels",
  "dispatch2",
+ "fancy-regex 0.14.0",
  "float8",
  "futures",
  "half",
  "hf-hub",
+ "indexmap 2.14.2",
  "indicatif 0.18.6",
  "lazy_static",
  "libc",
@@ -4057,9 +4064,9 @@ dependencies = [
 
 [[package]]
 name = "onde-mistralrs-sandbox"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "77845cc6e16e3b1eebed11d3ba6a3a56785f0031421bf46d1deefbe281aaa19e"
+checksum = "93b8ba017e502f2b7f2da67ced422ae73a7ea2494759b8c7a8e5fa38b341b3f5"
 dependencies = [
  "anyhow",
  "itoa",
@@ -4075,9 +4082,9 @@ dependencies = [
 
 [[package]]
 name = "onde-mistralrs-vision"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8cdc7ae244c416860838e9cb3ff5cd8ae8acf474f48f2fe15d45bc5491215760"
+checksum = "f231953141403eb6a825d6c247dffd7aa2e99b01b7a70d075bb72533c8dc10e8"
 dependencies = [
  "image",
  "onde-candle-core",
@@ -4096,17 +4103,6 @@ dependencies = [
  "tokio",
 ]
 
-[[package]]
-name = "onde-tqdm"
-version = "0.8.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2d96f4fcb4af06959bf9c9b81ace0502e2be2d11d7e037c94f3f7851bde720ea"
-dependencies = [
- "anyhow",
- "crossterm",
- "once_cell",
-]
-
 [[package]]
 name = "onde-uniffi-bindgen"
 version = "0.1.0"
@@ -4710,6 +4706,15 @@ dependencies = [
  "rand_core 0.10.1",
 ]
 
+[[package]]
+name = "rapidhash"
+version = "4.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5da7e78a036ce858e8d55b7e7dc8ba3a88b78350fd2155d3591bbd966b58589e"
+dependencies = [
+ "rustversion",
+]
+
 [[package]]
 name = "rav1e"
 version = "0.8.1"
@@ -6666,6 +6671,17 @@ version = "0.3.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8df9b6e13f2d32c91b9bd719c00d1958837bc7dec474d94952798cc8e69eeec3"
 
+[[package]]
+name = "tqdm"
+version = "0.8.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a882329b482cea1d2762e9b9ff1ed0f0327a1ec3a9ba7b8e8c4e3635908a4ae5"
+dependencies = [
+ "anyhow",
+ "crossterm",
+ "once_cell",
+]
+
 [[package]]
 name = "tracing"
 version = "0.1.44"
diff --git a/Cargo.toml b/Cargo.toml
index a013a83..fc153f4 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -45,51 +45,51 @@ crate-type = ["lib", "cdylib", "staticlib"]
 # in the Android sandbox).
 [target.'cfg(target_os = "android")'.dependencies]
 hf-hub = { version = "0.4", features = ["tokio", "rustls-tls"], default-features = false }
-mistralrs = { version = "0.9.3", package = "onde-mistralrs", features = [] }
-mistralrs-core = { version = "0.9.3", package = "onde-mistralrs-core", features = [] }
+mistralrs = { version = "0.9.5", package = "onde-mistralrs", features = [] }
+mistralrs-core = { version = "0.9.5", package = "onde-mistralrs-core", features = [] }
 
 # iOS: Metal-accelerated inference; hf-hub + mistralrs-core needed for
 # GLOBAL_HF_CACHE sandbox workaround (~/.cache is outside the container).
 [target.'cfg(target_os = "ios")'.dependencies]
 hf-hub = { version = "0.4", features = ["tokio", "rustls-tls"], default-features = false }
-mistralrs = { version = "0.9.3", package = "onde-mistralrs", features = ["metal"] }
-mistralrs-core = { version = "0.9.3", package = "onde-mistralrs-core", features = ["metal"] }
+mistralrs = { version = "0.9.5", package = "onde-mistralrs", features = ["metal"] }
+mistralrs-core = { version = "0.9.5", package = "onde-mistralrs-core", features = ["metal"] }
 
 # Linux: CPU-only baseline (no GPU features).
 # For NVIDIA GPU builds, patch this to features = ["cuda"] in CI before building.
 [target.'cfg(target_os = "linux")'.dependencies]
-mistralrs = { version = "0.9.3", package = "onde-mistralrs" }
+mistralrs = { version = "0.9.5", package = "onde-mistralrs" }
 
 # macOS: Metal-accelerated inference.
 [target.'cfg(target_os = "macos")'.dependencies]
-mistralrs = { version = "0.9.3", package = "onde-mistralrs", features = ["metal"] }
+mistralrs = { version = "0.9.5", package = "onde-mistralrs", features = ["metal"] }
 
 # tvOS: Metal-accelerated inference (Apple TV has Apple GPU); hf-hub +
 # mistralrs-core needed for GLOBAL_HF_CACHE sandbox workaround.
 [target.'cfg(target_os = "tvos")'.dependencies]
 hf-hub = { version = "0.4", features = ["tokio", "rustls-tls"], default-features = false }
-mistralrs = { version = "0.9.3", package = "onde-mistralrs", features = ["metal"] }
-mistralrs-core = { version = "0.9.3", package = "onde-mistralrs-core", features = ["metal"] }
+mistralrs = { version = "0.9.5", package = "onde-mistralrs", features = ["metal"] }
+mistralrs-core = { version = "0.9.5", package = "onde-mistralrs-core", features = ["metal"] }
 
 # visionOS: Metal-accelerated inference (Apple Vision Pro has M2 GPU);
 # hf-hub + mistralrs-core needed for GLOBAL_HF_CACHE sandbox workaround.
 [target.'cfg(target_os = "visionos")'.dependencies]
 hf-hub = { version = "0.4", features = ["tokio", "rustls-tls"], default-features = false }
-mistralrs = { version = "0.9.3", package = "onde-mistralrs", features = ["metal"] }
-mistralrs-core = { version = "0.9.3", package = "onde-mistralrs-core", features = ["metal"] }
+mistralrs = { version = "0.9.5", package = "onde-mistralrs", features = ["metal"] }
+mistralrs-core = { version = "0.9.5", package = "onde-mistralrs-core", features = ["metal"] }
 
 # watchOS: Metal-accelerated inference (Apple Watch Series 4+ has Apple GPU);
 # hf-hub + mistralrs-core needed for GLOBAL_HF_CACHE sandbox workaround.
 # Memory constrained — default to 1.5B model.
 [target.'cfg(target_os = "watchos")'.dependencies]
 hf-hub = { version = "0.4", features = ["tokio", "rustls-tls"], default-features = false }
-mistralrs = { version = "0.9.3", package = "onde-mistralrs", features = ["metal"] }
-mistralrs-core = { version = "0.9.3", package = "onde-mistralrs-core", features = ["metal"] }
+mistralrs = { version = "0.9.5", package = "onde-mistralrs", features = ["metal"] }
+mistralrs-core = { version = "0.9.5", package = "onde-mistralrs-core", features = ["metal"] }
 
 # Windows: CPU-only baseline (no GPU features).
 # For NVIDIA GPU builds, patch this to features = ["cuda"] in CI before building.
 [target.'cfg(target_os = "windows")'.dependencies]
-mistralrs = { version = "0.9.3", package = "onde-mistralrs" }
+mistralrs = { version = "0.9.5", package = "onde-mistralrs" }
 
 [dependencies]
 anyhow = "1"
@@ -119,14 +119,16 @@ dotenvy = "0.15"
 uniffi = { workspace = true, features = ["build"] }
 
 # The target-conditional mistralrs deps above resolve to the published
-# onde-mistralrs{,-core} 0.9.3 crates (setoelkahfi/mistral.rs fork,
-# fix/all-platform-fixes, synced with upstream 0.9.x). Those in turn pull
+# onde-mistralrs{,-core} 0.9.5 crates (setoelkahfi/mistral.rs fork,
+# fix/all-platform-fixes, synced with upstream 0.9.5). Those in turn pull
 # onde-candle-{core,nn} 0.11.0 (candle rev 27f20fea published under new names;
 # upstream mistral.rs needs post-0.11.0 candle APIs no released candle has).
-# For local dev against the fork, point them back at
-# git = "https://github.com/setoelkahfi/mistral.rs", branch = "fix/all-platform-fixes"
-# with package = "onde-mistralrs{,-core}" (the fork's crates carry the onde names
-# now); flip to registry form before publishing, cargo publish rejects git deps.
+# To test unpublished fork changes without editing this file, patch by the real
+# package name from the command line (a [patch.crates-io] rename is silently
+# ignored, but a patch keyed on the onde-* name works):
+#   cargo check --config 'patch.crates-io.onde-mistralrs.git="https://github.com/setoelkahfi/mistral.rs"' \
+#               --config 'patch.crates-io.onde-mistralrs.branch="fix/all-platform-fixes"'
+# and the same for onde-mistralrs-core. Do not commit the Cargo.lock that results.
 
 # Patch sysctl with watchOS support until upstream merges it.
 # PR: https://github.com/johalun/sysctl-rs/pull/74
diff --git a/src/hf_cache.rs b/src/hf_cache.rs
index fa71ef3..edb703e 100644
--- a/src/hf_cache.rs
+++ b/src/hf_cache.rs
@@ -1084,6 +1084,18 @@ where
         builder = builder.with_chat_template(chat_template);
     }
 
+    // Upstream #2493 made Metal paged attention the default; skip it on mobile
+    // so the validation load does not reserve the KV block pool.
+    #[cfg(any(
+        target_os = "ios",
+        target_os = "tvos",
+        target_os = "visionos",
+        target_os = "watchos"
+    ))]
+    {
+        builder = builder.without_paged_attn();
+    }
+
     // Shared flag so the monitor task knows when to stop.
     let finished = Arc::new(AtomicBool::new(false));
     let finished_clone = Arc::clone(&finished);
diff --git a/src/inference/engine.rs b/src/inference/engine.rs
index 5299e32..e82c554 100644
--- a/src/inference/engine.rs
+++ b/src/inference/engine.rs
@@ -472,17 +472,23 @@ impl ChatEngine {
         // so the auto device mapper allocates far less memory.
         #[cfg(target_os = "tvos")]
         {
-            use mistralrs::DeviceMapSetting;
-            use mistralrs_core::AutoDeviceMapParams;
-
             builder = builder
                 .with_max_num_seqs(1)
                 .with_prefix_cache_n(None)
                 .with_no_kv_cache()
-                .with_device_mapping(DeviceMapSetting::Auto(AutoDeviceMapParams::Text {
-                    max_seq_len: 512,
-                    max_batch_size: 1,
-                }));
+                .with_max_model_len(512);
+        }
+
+        // Upstream #2493 made Metal paged attention the default; its KV block
+        // pool is too large for the memory budget of Apple's mobile targets.
+        #[cfg(any(
+            target_os = "ios",
+            target_os = "tvos",
+            target_os = "visionos",
+            target_os = "watchos"
+        ))]
+        {
+            builder = builder.without_paged_attn();
         }
 
         // On Android the GGUF embedded tokenizer is not supported by the
@@ -638,17 +644,23 @@ impl ChatEngine {
 
         #[cfg(target_os = "tvos")]
         {
-            use mistralrs::DeviceMapSetting;
-            use mistralrs_core::AutoDeviceMapParams;
-
             builder = builder
                 .with_max_num_seqs(1)
                 .with_prefix_cache_n(None)
                 .with_no_kv_cache()
-                .with_device_mapping(DeviceMapSetting::Auto(AutoDeviceMapParams::Text {
-                    max_seq_len: 512,
-                    max_batch_size: 1,
-                }));
+                .with_max_model_len(512);
+        }
+
+        // Upstream #2493 made Metal paged attention the default; its KV block
+        // pool is too large for the memory budget of Apple's mobile targets.
+        #[cfg(any(
+            target_os = "ios",
+            target_os = "tvos",
+            target_os = "visionos",
+            target_os = "watchos"
+        ))]
+        {
+            builder = builder.without_paged_attn();
         }
 
         if let Some(ref template) = config.chat_template {

From 77523af173f0a42836f4ad52c114d8a39b5d25bf Mon Sep 17 00:00:00 2001
From: Seto Elkahfi <1797197+setoelkahfi@users.noreply.github.com>
Date: Fri, 30 Oct 2026 14:05:41 +0100
Subject: [PATCH 2/3] 1.3.4

Bump the root crate, the Kotlin, Dart and React Native packages, and the
root and per-SDK lockfiles from 1.3.3 to 1.3.4. Add changelog entries for
the onde-mistralrs 0.9.5 move and the mobile paged-attention change.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---
 CHANGELOG.md                     | 13 +++++++++++++
 Cargo.lock                       |  4 ++--
 Cargo.toml                       |  2 +-
 sdk/dart/CHANGELOG.md            |  5 +++++
 sdk/dart/pubspec.yaml            |  2 +-
 sdk/dart/rust/Cargo.lock         |  2 +-
 sdk/kotlin/gradle.properties     |  2 +-
 sdk/react-native/CHANGELOG.md    |  5 +++++
 sdk/react-native/package.json    |  2 +-
 sdk/react-native/rust/Cargo.lock |  4 ++--
 sdk/react-native/rust/Cargo.toml |  2 +-
 11 files changed, 33 insertions(+), 10 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 6545fb7..09acfaa 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,3 +1,16 @@
+## 1.3.4
+
+### Inference engine
+
+* Moves `onde-mistralrs` and `onde-mistralrs-core` from `0.9.3` to `0.9.5` on every target. The new crates merge upstream mistral.rs `0.9.5` into the fork and add visionOS and watchOS arms to the Metal `embedded_library!` macro, so `--features metal` builds for tvOS, visionOS and watchOS again.
+* Upstream made Metal paged attention the default. Its KV block pool does not fit in memory on iOS, tvOS, visionOS or watchOS, so model loading (and the validation load in `hf_cache`) now calls `without_paged_attn()` on those targets. Desktop and Android are unchanged.
+* On tvOS the auto device map override is replaced by `with_max_model_len(512)`.
+
+### Packaging
+
+* No public API changes: the published `onde` crate surface is identical to `1.3.3`.
+* All version sources (root, Kotlin, Dart, React Native, and the root and per-SDK lockfiles) are aligned on `1.3.4`.
+
 ## 1.3.3
 
 ### Cargo workspace
diff --git a/Cargo.lock b/Cargo.lock
index 2e86816..a96d33d 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -3729,7 +3729,7 @@ checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
 
 [[package]]
 name = "onde"
-version = "1.3.3"
+version = "1.3.4"
 dependencies = [
  "anyhow",
  "cc",
@@ -4093,7 +4093,7 @@ dependencies = [
 
 [[package]]
 name = "onde-react-native"
-version = "1.3.3"
+version = "1.3.4"
 dependencies = [
  "jni 0.21.1",
  "log",
diff --git a/Cargo.toml b/Cargo.toml
index fc153f4..d40db41 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "onde"
-version = "1.3.3"
+version = "1.3.4"
 edition = "2021"
 description = "On-device inference engine for Apple silicon."
 license = "MIT OR Apache-2.0"
diff --git a/sdk/dart/CHANGELOG.md b/sdk/dart/CHANGELOG.md
index 6ac3594..4d367f0 100644
--- a/sdk/dart/CHANGELOG.md
+++ b/sdk/dart/CHANGELOG.md
@@ -1,3 +1,8 @@
+## 1.3.4
+
+- Version aligned with the Rust core. No Dart API changes.
+- The core now builds on `onde-mistralrs` `0.9.5`. Paged attention is turned off on iOS, tvOS, visionOS and watchOS to stay inside the memory budget there.
+
 ## 1.3.3
 
 - Version aligned with the Rust core. No Dart API changes.
diff --git a/sdk/dart/pubspec.yaml b/sdk/dart/pubspec.yaml
index b6b8d82..0125cf8 100644
--- a/sdk/dart/pubspec.yaml
+++ b/sdk/dart/pubspec.yaml
@@ -2,7 +2,7 @@ name: onde_inference
 description: >-
   On-device LLM inference for Flutter & Dart. Run Qwen 2.5 models locally
   with Metal on iOS and macOS, CPU on Android and desktop. No cloud, no API key.
-version: 1.3.3
+version: 1.3.4
 repository: https://github.com/ondeinference/onde
 homepage: https://ondeinference.com
 documentation: https://ondeinference.com/sdk/dart
diff --git a/sdk/dart/rust/Cargo.lock b/sdk/dart/rust/Cargo.lock
index 9b27ac5..9b2df99 100644
--- a/sdk/dart/rust/Cargo.lock
+++ b/sdk/dart/rust/Cargo.lock
@@ -3604,7 +3604,7 @@ checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
 
 [[package]]
 name = "onde"
-version = "1.3.3"
+version = "1.3.4"
 dependencies = [
  "anyhow",
  "cc",
diff --git a/sdk/kotlin/gradle.properties b/sdk/kotlin/gradle.properties
index 0e4d26b..ed56f80 100644
--- a/sdk/kotlin/gradle.properties
+++ b/sdk/kotlin/gradle.properties
@@ -10,7 +10,7 @@ kotlin.code.style=official
 # Maven publishing coordinates — keep in sync with Cargo.toml [package] version
 GROUP=com.ondeinference
 POM_ARTIFACT_ID=onde-inference
-VERSION_NAME=1.3.3
+VERSION_NAME=1.3.4
 
 
 
diff --git a/sdk/react-native/CHANGELOG.md b/sdk/react-native/CHANGELOG.md
index b229132..d7931bb 100644
--- a/sdk/react-native/CHANGELOG.md
+++ b/sdk/react-native/CHANGELOG.md
@@ -1,3 +1,8 @@
+## 1.3.4
+
+- Version aligned with the Rust core. No React Native API changes.
+- The core now builds on `onde-mistralrs` `0.9.5`. Paged attention is turned off on iOS, tvOS, visionOS and watchOS to stay inside the memory budget there.
+
 ## 1.3.3
 
 - Version aligned with the Rust core. No React Native API changes.
diff --git a/sdk/react-native/package.json b/sdk/react-native/package.json
index 76bc203..7e1c00d 100644
--- a/sdk/react-native/package.json
+++ b/sdk/react-native/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@ondeinference/react-native",
-  "version": "1.3.3",
+  "version": "1.3.4",
   "description": "On-device LLM inference for React Native. Run Qwen 2.5 models locally with Metal on iOS, CPU on Android. No cloud, no API key.",
   "main": "build/index.js",
   "types": "build/index.d.ts",
diff --git a/sdk/react-native/rust/Cargo.lock b/sdk/react-native/rust/Cargo.lock
index 06780f1..999eaa7 100644
--- a/sdk/react-native/rust/Cargo.lock
+++ b/sdk/react-native/rust/Cargo.lock
@@ -3420,7 +3420,7 @@ checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
 
 [[package]]
 name = "onde"
-version = "1.3.3"
+version = "1.3.4"
 dependencies = [
  "anyhow",
  "cc",
@@ -3777,7 +3777,7 @@ dependencies = [
 
 [[package]]
 name = "onde-react-native"
-version = "1.3.3"
+version = "1.3.4"
 dependencies = [
  "jni",
  "log",
diff --git a/sdk/react-native/rust/Cargo.toml b/sdk/react-native/rust/Cargo.toml
index f959987..04c786c 100644
--- a/sdk/react-native/rust/Cargo.toml
+++ b/sdk/react-native/rust/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "onde-react-native"
-version = "1.3.3"
+version = "1.3.4"
 edition = "2021"
 publish = false
 

From 3f2bc565644b99a2b2d6b5deda524ff4ad60f05f Mon Sep 17 00:00:00 2001
From: Seto Elkahfi <1797197+setoelkahfi@users.noreply.github.com>
Date: Fri, 30 Oct 2026 15:14:07 +0100
Subject: [PATCH 3/3] Sync the React Native package-lock to 1.3.4

The release preflight runs make verify, which checks that
sdk/react-native/package-lock.json agrees with Cargo.toml. The 1.3.4
bump edited the other manifests by hand and missed it, so preflight
failed on both version fields. Running scripts/sync-release-version.py
updates the lock and nothing else.

Co-Authored-By: OndeCode v1.0.3-acp <noreply@ondeinference.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---
 sdk/react-native/package-lock.json | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/sdk/react-native/package-lock.json b/sdk/react-native/package-lock.json
index c91185c..8631279 100644
--- a/sdk/react-native/package-lock.json
+++ b/sdk/react-native/package-lock.json
@@ -1,12 +1,12 @@
 {
   "name": "@ondeinference/react-native",
-  "version": "1.3.3",
+  "version": "1.3.4",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@ondeinference/react-native",
-      "version": "1.3.3",
+      "version": "1.3.4",
       "license": "MIT OR Apache-2.0",
       "devDependencies": {
         "expo-modules-core": "^2.0.0",