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/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 eb13a5d..a96d33d 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",
@@ -3729,7 +3729,7 @@ checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
 
 [[package]]
 name = "onde"
-version = "1.3.3"
+version = "1.3.4"
 dependencies = [
  "anyhow",
  "cc",
@@ -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",
@@ -4086,7 +4093,7 @@ dependencies = [
 
 [[package]]
 name = "onde-react-native"
-version = "1.3.3"
+version = "1.3.4"
 dependencies = [
  "jni 0.21.1",
  "log",
@@ -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..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"
@@ -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/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-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",
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
 
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 {