Skip to main content

mlx_native/
kernel_registry.rs

1//! [`KernelRegistry`] — lazy compilation and caching of Metal compute pipelines.
2//!
3//! MSL shader source is embedded at compile time via `include_str!`.  On first
4//! access, the source is compiled into a Metal library, the named function is
5//! extracted, and a `ComputePipelineState` is created and cached.  Subsequent
6//! calls return the cached pipeline.
7//!
8//! ## ADR-029 iter-175 Step 1l: precompiled `.metallib` fast path
9//!
10//! `build.rs` runs `xcrun metal -O3` on every `.metal` file under
11//! `src/shaders/` and links the results into a single `default.metallib`
12//! placed in `OUT_DIR`.  We embed the bytes via `include_bytes!`.
13//!
14//! When `MLX_PRECOMPILED_METALLIB=1` is set, `get_pipeline` and
15//! `get_pipeline_with_constants` first try to resolve the kernel function
16//! against this precompiled library; if found, build the pipeline from it
17//! (saves Apple's runtime source-compile pass).  On any failure (function
18//! missing, empty embedded blob, load error) the code transparently falls
19//! back to the original source-compile path — byte-identical behavior.
20//!
21//! Empirical motivation (iter 1k test
22//! `tests/iter175_h_e_metallib_perf.rs`): on M5 Max at the gemma4 Q_sliding
23//! decode shape, precompiled is **+5.89% faster** than runtime-source
24//! (median 18.35 µs vs 19.49 µs) with a tighter p90 (19.54 vs 70.11) —
25//! Apple's runtime compile chain has occasional jitter outliers.
26
27use std::collections::HashMap;
28use std::sync::OnceLock;
29
30use metal::{ComputePipelineDescriptor, ComputePipelineState, FunctionConstantValues, MTLDataType};
31
32use crate::error::{MlxError, Result};
33
34/// Bytes of the precompiled `default.metallib` produced by `build.rs` from
35/// every `src/shaders/*.metal` file.  Empty when `MLX_NATIVE_SKIP_METALLIB`
36/// was set at build time or xcrun was unavailable.
37const EMBEDDED_METALLIB: &[u8] = include_bytes!(concat!(env!("OUT_DIR"), "/default.metallib"));
38
39/// Returns `true` when the precompiled `.metallib` fast path is enabled
40/// (default-ON since Step 1m validation).  Set `MLX_PRECOMPILED_METALLIB=0`
41/// (or `false`, `off`) to opt out — useful for diagnosing kernel-compile
42/// regressions or A/B benching.
43///
44/// Step 1m multi-regime validation at HEAD (gemma4-APEX-Q5_K_M, M5 Max):
45/// - tg100 alt-pair 2-cycle: decode 95.5 → 95.9 t/s (+0.42%), prefill
46///   179 → 185 t/s (+3.3%)
47/// - tg2000 alt-pair 1-cycle: decode 93.1 → 92.8 t/s (−0.32%, within σ),
48///   prefill 171 → 178.5 t/s (+4.4%)
49/// - coherence_smoke: 2/2 PASS both configs
50///
51/// The Step 1l initial smoke (`tg50 95.4 → 62.1`) did NOT reproduce on
52/// rebuild — most likely a cold-PSO-cache or first-run-of-rebuilt-binary
53/// artifact rather than a real perf regression.
54fn precompiled_enabled() -> bool {
55    static FLAG: OnceLock<bool> = OnceLock::new();
56    *FLAG.get_or_init(|| {
57        match std::env::var("MLX_PRECOMPILED_METALLIB").as_deref() {
58            Ok("0") | Ok("false") | Ok("off") => false,
59            _ => true,
60        }
61    })
62}
63
64/// Returns `true` when the precompiled `.metallib` is consulted for
65/// `get_pipeline_with_constants` (FCV-specialized) kernels.  Inherits
66/// the master gate [`precompiled_enabled`]; both must be ON for the
67/// FCV path to use precompiled.  Default-ON since Step 1m validation.
68fn precompiled_fcv_enabled() -> bool {
69    static FLAG: OnceLock<bool> = OnceLock::new();
70    *FLAG.get_or_init(|| {
71        match std::env::var("MLX_PRECOMPILED_METALLIB_FCV").as_deref() {
72            Ok("0") | Ok("false") | Ok("off") => false,
73            _ => true,
74        }
75    })
76}
77
78// MTLDataType numeric values (from metal-rs argument.rs, confirmed in Apple Metal spec):
79//   Int  = 29
80//   Bool = 53
81// These are used when calling set_constant_value_at_index so the Metal runtime
82// knows how wide each constant value is.
83
84/// Registry that lazily compiles and caches Metal compute pipelines from
85/// embedded MSL source.
86///
87/// # Usage
88///
89/// ```ignore
90/// let mut registry = KernelRegistry::new();
91/// let pipeline = registry.get_pipeline("elementwise_add", device.metal_device())?;
92/// encoder.encode(&pipeline, &buffers, grid, tg);
93/// ```
94///
95/// # Thread Safety
96///
97/// `KernelRegistry` is **not** `Sync` by default (it uses `&mut self` for
98/// `get_pipeline` to allow mutable cache insertion).  If you need concurrent
99/// access, wrap it in a `Mutex` or use one registry per thread.
100pub struct KernelRegistry {
101    /// Cached pipelines keyed by kernel function name.
102    cache: HashMap<String, ComputePipelineState>,
103    /// MSL source text keyed by kernel function name.
104    ///
105    /// Populated at construction time with all embedded shader sources.
106    sources: HashMap<String, &'static str>,
107    /// ADR-029 iter-175 Step 1l: precompiled `default.metallib` (lazy).
108    ///
109    /// `None` initially.  On first `get_pipeline*` call under
110    /// `MLX_PRECOMPILED_METALLIB=1`, populated via
111    /// `device.new_library_with_data(EMBEDDED_METALLIB)` — or set to a
112    /// sentinel "load-failed" marker (still `None`) so we don't retry.
113    /// Set to `Some(library)` on success.
114    ///
115    /// Lazily filled because we need a `&metal::DeviceRef` to load it
116    /// and `KernelRegistry::new()` does not have one.
117    precompiled_lib: Option<metal::Library>,
118    /// Whether we've already attempted to load the precompiled library.
119    /// Prevents repeated load attempts on failure.
120    precompiled_load_attempted: bool,
121}
122
123impl KernelRegistry {
124    /// Create a new registry with all embedded shader sources pre-registered.
125    ///
126    /// No compilation happens here — shaders are compiled lazily on first use.
127    pub fn new() -> Self {
128        let mut sources = HashMap::new();
129
130        // Register embedded shader sources.
131        sources.insert(
132            "placeholder".into(),
133            include_str!("shaders/placeholder.metal"),
134        );
135        sources.insert(
136            "quantized_matmul".into(),
137            include_str!("shaders/quantized_matmul.metal"),
138        );
139        sources.insert(
140            "quantized_matmul_simd".into(),
141            include_str!("shaders/quantized_matmul.metal"),
142        );
143        sources.insert(
144            "quantized_matmul_simd_bf16".into(),
145            include_str!("shaders/quantized_matmul.metal"),
146        );
147        sources.insert(
148            "quantized_matmul_simd_bf16_expert".into(),
149            include_str!("shaders/quantized_matmul.metal"),
150        );
151
152        // GGML block-format quantized mat-vec kernels (ADR-006 Phase 3)
153        let ggml_src: &'static str =
154            include_str!("shaders/quantized_matmul_ggml.metal");
155        sources.insert("kernel_mul_mv_q4_0_f32".into(), ggml_src);
156        sources.insert("kernel_mul_mv_q8_0_f32".into(), ggml_src);
157        // ADR-028 iter-368: peer-style NSG=4 NR=2 variant (128 threads/TG).
158        sources.insert("kernel_mul_mv_q8_0_f32_nr2".into(), ggml_src);
159        sources.insert("kernel_mul_mv_q6_K_f32".into(), ggml_src);
160        // ADR-028 iter-309 — q6_K mat-vec with nr0=2 + cached yl[16]
161        // (peer-pattern port of llama.cpp's `kernel_mul_mv_q6_K_f32_impl`
162        // with N_R0_Q6_K=2; 4 rows/TG vs baseline's 2).  Env-gated via
163        // `HF2Q_Q6K_MV_NR2=1` in the dispatcher.
164        sources.insert("kernel_mul_mv_q6_K_f32_nr2".into(), ggml_src);
165        // ADR-022 Phase 1 — Q5_1 / IQ4_NL dense mat-vec.
166        sources.insert("kernel_mul_mv_q5_1_f32".into(), ggml_src);
167        sources.insert("kernel_mul_mv_iq4_nl_f32".into(), ggml_src);
168        // ADR-013 P7 — Q4_K dense decode mat-vec (port of llama.cpp's
169        // kernel_mul_mv_q4_K_f32 at ggml-metal.metal:7715-7821).
170        sources.insert("kernel_mul_mv_q4_K_f32".into(), ggml_src);
171        // ADR-022 Phase 2 — Q5_K dense mv kernel.
172        sources.insert("kernel_mul_mv_q5_K_f32".into(), ggml_src);
173
174        // GGML block-format quantized matrix-matrix kernels
175        // (ADR-011 Phase 3 Wave P3a: port of llama.cpp's kernel_mul_mm_<q>_f32).
176        // Used at prefill m > 8 to reuse each weight tile across a 32-row
177        // block via threadgroup-staged simdgroup MMA, instead of re-reading
178        // every block per prompt-token as the mv kernel does.
179        let ggml_mm_src: &'static str =
180            include_str!("shaders/quantized_matmul_mm.metal");
181        sources.insert("kernel_mul_mm_q4_0_f32".into(), ggml_mm_src);
182        sources.insert("kernel_mul_mm_q8_0_f32".into(), ggml_mm_src);
183        sources.insert("kernel_mul_mm_q6_K_f32".into(), ggml_mm_src);
184        // ADR-022 Phase 1 — dense Q5_1 / IQ4_NL mm.
185        sources.insert("kernel_mul_mm_q5_1_f32".into(), ggml_mm_src);
186        sources.insert("kernel_mul_mm_iq4_nl_f32".into(), ggml_mm_src);
187        // ADR-022 Phase 2 — dense Q5_K mm.
188        sources.insert("kernel_mul_mm_q5_K_f32".into(), ggml_mm_src);
189        // ADR-022 Phase 3 — dense Q4_K mm.
190        sources.insert("kernel_mul_mm_q4_K_f32".into(), ggml_mm_src);
191
192        // GGML block-format quantized matrix-matrix kernels — tensor API
193        // variant (ADR-011 Phase 3 Wave P3b-tensor: port of llama.cpp's
194        // kernel_mul_mm_impl `#ifdef GGML_METAL_HAS_TENSOR` branch).
195        // Uses Apple's MetalPerformancePrimitives `tensor_ops::matmul2d`
196        // primitive which on M3+ dispatches to hardware tensor cores for
197        // 2-3x the effective FLOP throughput vs the simdgroup MMA path.
198        // Only compiled on devices where the tensor API is available; the
199        // kernel_registry's runtime-probe (see MlxDevice::has_tensor) gates
200        // compilation so non-tensor devices transparently fall back to the
201        // non-tensor `kernel_mul_mm_<q>_f32` kernels.
202        let ggml_mm_tensor_src: &'static str =
203            include_str!("shaders/quantized_matmul_mm_tensor.metal");
204        sources.insert("kernel_mul_mm_q4_0_tensor_f32".into(), ggml_mm_tensor_src);
205        sources.insert("kernel_mul_mm_q4_0_tensor_bf16_perm021".into(), ggml_mm_tensor_src);
206        sources.insert("kernel_mul_mm_q6_K_tensor_bf16_perm021".into(), ggml_mm_tensor_src);
207        sources.insert("kernel_mul_mm_q8_0_tensor_f32".into(), ggml_mm_tensor_src);
208        sources.insert("kernel_mul_mm_q6_K_tensor_f32".into(), ggml_mm_tensor_src);
209        // ADR-022 Phase 1 — Q5_1 / IQ4_NL tensor mm.
210        sources.insert("kernel_mul_mm_q5_1_tensor_f32".into(), ggml_mm_tensor_src);
211        sources.insert("kernel_mul_mm_iq4_nl_tensor_f32".into(), ggml_mm_tensor_src);
212        // ADR-022 Phase 2 — Q5_K tensor mm.
213        sources.insert("kernel_mul_mm_q5_K_tensor_f32".into(), ggml_mm_tensor_src);
214        // ADR-022 Phase 3 — Q4_K tensor mm + Q8_0 perm021.
215        sources.insert("kernel_mul_mm_q4_K_tensor_f32".into(), ggml_mm_tensor_src);
216        sources.insert("kernel_mul_mm_q8_0_tensor_bf16_perm021".into(), ggml_mm_tensor_src);
217        // ADR-029 iter-30 H29-speed — F16-weight V2 large-tile mm.
218        // Same source file as the V2 quantized variants; reads F16 weight
219        // directly from device memory (no per-call dequant).  Used when
220        // MlxQWeight.f16_shadow is populated and m > MM_ROUTING_THRESHOLD.
221        sources.insert("hf2q_mul_mm_tensor_v2_f16".into(), ggml_mm_tensor_src);
222        // ADR-029 iter-36 H28-D — F16-weight perm021 mm for O-projection.
223        // Same source file; reads F16 weight from MlxQWeight.f16_shadow when
224        // populated, bypassing the per-call quantized dequant.  B-stage
225        // (bfloat permuted [n_heads, seq_len, head_dim] input) is byte-
226        // identical to the quantized variant.
227        sources.insert("kernel_mul_mm_f16_tensor_bf16_perm021".into(), ggml_mm_tensor_src);
228        // ADR-029 iter-23 H28-A — V2 large-tile tensor mm (NRA=64 M, NRB=128 N).
229        // Same source file as V1 tensor mm; distinct kernel host names so the
230        // dispatcher can pick V1 vs V2 at runtime via HF2Q_LARGE_TILE_MM.
231        sources.insert("kernel_mul_mm_q4_0_tensor_v2_f32".into(), ggml_mm_tensor_src);
232        sources.insert("kernel_mul_mm_q8_0_tensor_v2_f32".into(), ggml_mm_tensor_src);
233        sources.insert("kernel_mul_mm_q6_K_tensor_v2_f32".into(), ggml_mm_tensor_src);
234        sources.insert("kernel_mul_mm_q5_1_tensor_v2_f32".into(), ggml_mm_tensor_src);
235        sources.insert("kernel_mul_mm_iq4_nl_tensor_v2_f32".into(), ggml_mm_tensor_src);
236        sources.insert("kernel_mul_mm_q5_K_tensor_v2_f32".into(), ggml_mm_tensor_src);
237        sources.insert("kernel_mul_mm_q4_K_tensor_v2_f32".into(), ggml_mm_tensor_src);
238        // ADR-029 iter-28 H29 — whole-tensor dequant from block_q → F16.
239        // Used at model load to materialize an F16 shadow of attn/dense MLP
240        // weights so the runtime dispatch can use kernel_mul_mm_f16_f32_*
241        // (peer's gemma4 pattern).  Trades ~1 GB resident memory for 2-3×
242        // faster per-call dense matmul at prefill.
243        let dequant_to_f16_src: &'static str =
244            include_str!("shaders/dequant_to_f16.metal");
245        sources.insert("hf2q_dequant_q4_0_to_f16".into(), dequant_to_f16_src);
246        sources.insert("hf2q_dequant_q8_0_to_f16".into(), dequant_to_f16_src);
247        sources.insert("hf2q_dequant_q5_1_to_f16".into(), dequant_to_f16_src);
248        sources.insert("hf2q_dequant_iq4_nl_to_f16".into(), dequant_to_f16_src);
249        sources.insert("hf2q_dequant_q4_K_to_f16".into(), dequant_to_f16_src);
250        sources.insert("hf2q_dequant_q5_K_to_f16".into(), dequant_to_f16_src);
251        sources.insert("hf2q_dequant_q6_K_to_f16".into(), dequant_to_f16_src);
252
253        // ADR-022 Phase 1 P1.7 — Q5_1 / IQ4_NL mul_mv_ext r1 family.
254        // Eight instantiations (2 types × 4 r1ptg widths). Each PSO is
255        // additionally specialized at PSO-compile time with FC_mul_mv_nsg
256        // (function_constant 600) and FC_mul_mv_nxpsg (function_constant 601).
257        let mul_mv_ext_src: &'static str = include_str!("shaders/mul_mv_ext.metal");
258        sources.insert("kernel_mul_mv_ext_q5_1_f32_r1_2".into(), mul_mv_ext_src);
259        sources.insert("kernel_mul_mv_ext_q5_1_f32_r1_3".into(), mul_mv_ext_src);
260        sources.insert("kernel_mul_mv_ext_q5_1_f32_r1_4".into(), mul_mv_ext_src);
261        sources.insert("kernel_mul_mv_ext_q5_1_f32_r1_5".into(), mul_mv_ext_src);
262        sources.insert("kernel_mul_mv_ext_iq4_nl_f32_r1_2".into(), mul_mv_ext_src);
263        sources.insert("kernel_mul_mv_ext_iq4_nl_f32_r1_3".into(), mul_mv_ext_src);
264        sources.insert("kernel_mul_mv_ext_iq4_nl_f32_r1_4".into(), mul_mv_ext_src);
265        sources.insert("kernel_mul_mv_ext_iq4_nl_f32_r1_5".into(), mul_mv_ext_src);
266        // ADR-022 Phase 4 — Q4_0 / Q8_0 / Q4_K / Q5_K / Q6_K mv_ext.
267        // 5 types × 4 r1ptg widths = 20 instantiations.
268        for r1 in [2, 3, 4, 5].iter() {
269            for ty in ["q4_0", "q8_0", "q4_K", "q5_K", "q6_K"].iter() {
270                let name = format!("kernel_mul_mv_ext_{ty}_f32_r1_{r1}");
271                sources.insert(name, mul_mv_ext_src);
272            }
273        }
274
275        // Dense bf16×f32 → f32 tensor-API matmul (non-flash-attention
276        // prefill Q@K^T and scores@V, modeled on llama.cpp's
277        // kernel_mul_mm_bf16_f32 with the GGML_METAL_HAS_TENSOR branch
278        // active).  Tile geometry and write-back identical to the
279        // quantized tensor kernel; only the A-stage copy (bfloat →
280        // bfloat, no dequantize) differs.
281        let dense_mm_bf16_tensor_src: &'static str =
282            include_str!("shaders/dense_mm_bf16_tensor.metal");
283        sources.insert("hf2q_dense_mm_bf16_f32_tensor".into(), dense_mm_bf16_tensor_src);
284        // ADR-029 iter-80 H60: V2 large-tile variant (NRA=64, NRB=128).
285        // Same source file (`dense_mm_bf16_tensor.metal`) — second host_name
286        // entry resolves to the V2 kernel appended at the bottom of that
287        // file. Picked at dispatch time when HF2Q_LARGE_TILE_MM=1.
288        sources.insert("hf2q_dense_mm_bf16_f32_tensor_v2".into(), dense_mm_bf16_tensor_src);
289
290        // Dense f32×f32 → f32 tensor-API matmul (F32-everywhere
291        // sibling of dense_mm_bf16_tensor).  Used by hf2q's ADR-005
292        // iter-118 BF16-vs-F32 ViT attention A/B diagnostic to remove
293        // the BF16 K-stage cast as a confounding variable.  Port of
294        // llama.cpp's kernel_mul_mm_f32_f32 specialization
295        // (ggml-metal.metal:10098) on the GGML_METAL_HAS_TENSOR
296        // branch.  Same tile geometry (NR0=64 NR1=32 NK=32) but
297        // float-everywhere shmem staging.
298        let dense_mm_f32_f32_tensor_src: &'static str =
299            include_str!("shaders/dense_mm_f32_f32.metal");
300        sources.insert("hf2q_dense_mm_f32_f32_tensor".into(), dense_mm_f32_f32_tensor_src);
301
302        // Dense f16×f32 → f32 tensor-API matmul (F16-staging sibling
303        // of dense_mm_bf16_tensor).  Used by hf2q's ADR-005 Phase 2c
304        // iter-128 gemma4v ViT precision-parity path: every mmproj
305        // weight is stored as F16 in GGUF, peer's `kernel_mul_mm_f16_f32`
306        // (`ggml-metal.metal:10099`) stages BOTH A and B as `half` in
307        // shmem and computes on `simdgroup_half8x8`.  Matches peer
308        // per-element rounding budget exactly (10-bit mantissa vs
309        // BF16's 7-bit), closing the 1.16x/block cascade compound that
310        // iter-127 numerically bisected to BF16 staging.  Same tile
311        // geometry as the BF16 sibling (NR0=64 NR1=32 NK=32, 8 KB
312        // shmem) — half and bfloat share 16-bit storage.
313        let dense_mm_f16_tensor_src: &'static str =
314            include_str!("shaders/dense_mm_f16_tensor.metal");
315        sources.insert("hf2q_dense_mm_f16_f32_tensor".into(), dense_mm_f16_tensor_src);
316
317        // Dense bf16×f32 → f32 GEMV (matrix-vector multiply) — optimized
318        // for M=1 single-token decode.  Port of llama.cpp's
319        // kernel_mul_mv_bf16_f32_4 (bfloat4-vectorized GEMV kernel).
320        // Used in apply_linear_projection_f32 when seq_len=1 and the
321        // weight matrix is BF16, replacing the MM kernel (~2× faster for
322        // M=1 due to better memory bandwidth utilization per thread).
323        let dense_gemv_bf16_src: &'static str =
324            include_str!("shaders/dense_gemv_bf16.metal");
325        sources.insert("hf2q_dense_gemv_bf16_f32_4".into(), dense_gemv_bf16_src);
326
327        // Fused scale-mask-softmax for the non-flash-attention prefill
328        // path.  One row-local threadgroup per (head, query) pair
329        // replaces three separate dispatches (scale, mask-add, softmax);
330        // reads a bf16 mask (-INF at masked positions, matching
331        // flash_attn_prefill_mask.metal) that is shared across heads.
332        let scale_mask_softmax_src: &'static str =
333            include_str!("shaders/scale_mask_softmax.metal");
334        sources.insert("scale_mask_softmax_f32".into(), scale_mask_softmax_src);
335        // ADR-029 iter-93 H71: float4-vectorized variant for peer parity
336        // with kernel_soft_max_f32_4. Same source file; v4 host_name resolves
337        // to the second kernel appended at the bottom of scale_mask_softmax.metal.
338        sources.insert("scale_mask_softmax_f32_v4".into(), scale_mask_softmax_src);
339
340        // Expert-routed (MoE) quantized matmul kernel (Story 2.1)
341        sources.insert(
342            "quantized_matmul_id".into(),
343            include_str!("shaders/quantized_matmul_id.metal"),
344        );
345
346        // Expert-routed (MoE) GGML block-format quantized matmul kernels
347        let ggml_id_src: &'static str =
348            include_str!("shaders/quantized_matmul_id_ggml.metal");
349        sources.insert("kernel_mul_mv_id_q4_0_f32".into(), ggml_id_src);
350        sources.insert("kernel_mul_mv_id_q8_0_f32".into(), ggml_id_src);
351        // ADR-013 P7 — Q4_K MoE expert-routed mat-vec (port of
352        // llama.cpp's kernel_mul_mv_id_q4_K_f32 at ggml-metal.metal:10349).
353        sources.insert("kernel_mul_mv_id_q4_K_f32".into(), ggml_id_src);
354        sources.insert("kernel_mul_mv_id_q5_K_f32".into(), ggml_id_src);
355        sources.insert("kernel_mul_mv_id_q6_K_f32".into(), ggml_id_src);
356        // ADR-028 iter-321 — q6_K _id with nr0=2 + cached yl[16]
357        // (peer-pattern port mirroring iter-309's non-_id variant).
358        // Env-gated via HF2Q_Q6K_ID_MV_NR2=1 in dispatch_id_mv.
359        sources.insert("kernel_mul_mv_id_q6_K_f32_nr2".into(), ggml_id_src);
360        // ADR-029 iter-6 — q8_0 _id with nr0=2 + nsg=4 cross-SG reduce
361        // (peer-pattern port; peer N_R0_Q8_0=2 + N_SG_Q8_0=4 in
362        //  /opt/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h:27,40).
363        // Env-gated via HF2Q_Q8_0_ID_MV_NR2=1 in dispatch_id_mv.
364        sources.insert("kernel_mul_mv_id_q8_0_f32_nr2".into(), ggml_id_src);
365        // ADR-022 Phase 1 — Q5_1 / IQ4_NL MoE expert-routed mat-vec.
366        sources.insert("kernel_mul_mv_id_q5_1_f32".into(), ggml_id_src);
367        sources.insert("kernel_mul_mv_id_iq4_nl_f32".into(), ggml_id_src);
368        // Fused-SwiGLU mv_id variants (ADR-012 §Optimize / Task #15):
369        // computes y[r][n] = sum_k(dequant(W[expert][n][k]) * silu(gate[r][k]) * up[r][k])
370        // in one dispatch — replaces silu_mul + expert_down sequence.
371        sources.insert("kernel_mul_mv_id_q4_0_f32_swiglu".into(), ggml_id_src);
372
373        // Expert-routed (MoE) GGML block-format QUANTIZED MATRIX-MATRIX kernels
374        // (ADR-011 Phase 3 Wave P3a: port of llama.cpp's
375        // `kernel_mul_mm_id_map0_ne20_N` + `kernel_mul_mm_id_<q>_f32`).
376        // Two-stage dispatch: map0 regroups the token-to-expert table into
377        // per-expert routed-token lists, then mm_id stages a 64x32 expert
378        // weight tile into threadgroup shmem and reuses it across a 32-row
379        // block of that expert's routed tokens.
380        let ggml_id_mm_src: &'static str =
381            include_str!("shaders/quantized_matmul_id_mm.metal");
382        sources.insert("kernel_mul_mm_id_map0_ne20_1".into(), ggml_id_mm_src);
383        sources.insert("kernel_mul_mm_id_map0_ne20_8".into(), ggml_id_mm_src);
384        sources.insert("kernel_mul_mm_id_q4_0_f32".into(), ggml_id_mm_src);
385        sources.insert("kernel_mul_mm_id_q8_0_f32".into(), ggml_id_mm_src);
386        sources.insert("kernel_mul_mm_id_q6_K_f32".into(), ggml_id_mm_src);
387        // ADR-013 P16 — Q4_K mm_id (port of llama.cpp ggml-metal.metal:10169).
388        sources.insert("kernel_mul_mm_id_q4_K_f32".into(), ggml_id_mm_src);
389        // ADR-022 Phase 1 P1.6 — Q5_1 / IQ4_NL mm_id template instantiations.
390        sources.insert("kernel_mul_mm_id_q5_1_f32".into(), ggml_id_mm_src);
391        sources.insert("kernel_mul_mm_id_iq4_nl_f32".into(), ggml_id_mm_src);
392        // ADR-022 Phase 2 — Q5_K mm_id template instantiation.
393        sources.insert("kernel_mul_mm_id_q5_K_f32".into(), ggml_id_mm_src);
394
395        // MoE-routed quantized matrix-matrix kernels — tensor API variant
396        // (ADR-011 Phase 3 Wave P3b-tensor).  Uses the MPP tensor_ops
397        // matmul2d primitive for hardware-tensor-core MMA on M3+.  Only
398        // the mm_id kernel is ported — map0 is a short pre-pass (not
399        // matmul) and continues to use the simdgroup version.
400        let ggml_id_mm_tensor_src: &'static str =
401            include_str!("shaders/quantized_matmul_id_mm_tensor.metal");
402        sources.insert("kernel_mul_mm_id_q4_0_tensor_f32".into(), ggml_id_mm_tensor_src);
403        sources.insert("kernel_mul_mm_id_q8_0_tensor_f32".into(), ggml_id_mm_tensor_src);
404        sources.insert("kernel_mul_mm_id_q6_K_tensor_f32".into(), ggml_id_mm_tensor_src);
405        // ADR-013 P16 — Q4_K tensor-API mm_id.
406        sources.insert("kernel_mul_mm_id_q4_K_tensor_f32".into(), ggml_id_mm_tensor_src);
407        // ADR-022 Phase 1 P1.6 — Q5_1 / IQ4_NL tensor-API mm_id.
408        sources.insert("kernel_mul_mm_id_q5_1_tensor_f32".into(), ggml_id_mm_tensor_src);
409        sources.insert("kernel_mul_mm_id_iq4_nl_tensor_f32".into(), ggml_id_mm_tensor_src);
410        // ADR-022 Phase 2 — Q5_K tensor-API mm_id.
411        sources.insert("kernel_mul_mm_id_q5_K_tensor_f32".into(), ggml_id_mm_tensor_src);
412
413        // Embedding kernels (Story 1.5)
414        let embedding_src: &'static str = include_str!("shaders/embedding.metal");
415        sources.insert("embedding_gather_4bit".into(), embedding_src);
416        sources.insert("embedding_gather_6bit".into(), embedding_src);
417
418        // MoE gate kernel (Story 1.5)
419        let moe_gate_src: &'static str = include_str!("shaders/moe_gate.metal");
420        sources.insert("moe_gate".into(), moe_gate_src);
421
422        // MoE dispatch kernels (Story 1.5)
423        let moe_dispatch_src: &'static str = include_str!("shaders/moe_dispatch.metal");
424        sources.insert("fused_gelu_mul".into(), moe_dispatch_src);
425        sources.insert("moe_swiglu_fused".into(), moe_dispatch_src);
426        sources.insert("moe_swiglu_batch".into(), moe_dispatch_src);
427        sources.insert("moe_swiglu_seq".into(), moe_dispatch_src);
428        sources.insert("moe_accumulate".into(), moe_dispatch_src);
429        sources.insert("moe_weighted_sum".into(), moe_dispatch_src);
430        sources.insert("moe_weighted_sum_seq".into(), moe_dispatch_src);
431        sources.insert("zero_buffer".into(), moe_dispatch_src);
432        sources.insert("naive_matvec_f32".into(), moe_dispatch_src);
433        sources.insert("moe_gather_topk_weights".into(), moe_dispatch_src);
434        // bf16 variants (Phase 2 bf16 activation path)
435        sources.insert("fused_gelu_mul_bf16".into(), moe_dispatch_src);
436        sources.insert("moe_swiglu_seq_bf16".into(), moe_dispatch_src);
437        sources.insert("moe_weighted_sum_seq_bf16_input".into(), moe_dispatch_src);
438        // ADR-020 iter-11h-e3a: backward kernels for moe_weighted_sum_seq.
439        sources.insert(
440            "moe_weighted_sum_seq_backward_outputs_f32".into(),
441            moe_dispatch_src,
442        );
443        sources.insert(
444            "moe_weighted_sum_seq_backward_weights_f32".into(),
445            moe_dispatch_src,
446        );
447        // ADR-020 iter-11h-e3b: fused backward kernel for moe_swiglu_seq.
448        sources.insert(
449            "moe_swiglu_seq_backward_f32".into(),
450            moe_dispatch_src,
451        );
452
453        // Batched KV cache copy kernels
454        let kv_cache_src: &'static str = include_str!("shaders/kv_cache_copy.metal");
455        sources.insert("kv_cache_copy_batch_f32".into(), kv_cache_src);
456        sources.insert("kv_cache_copy_batch_f32_to_f16".into(), kv_cache_src);
457        sources.insert("kv_cache_copy_seq_f32".into(), kv_cache_src);
458        sources.insert("kv_cache_copy_seq_f32_to_f16".into(), kv_cache_src);
459        // Wave P4.11 — fused K+V copy variants
460        sources.insert("kv_cache_copy_seq_f32_kv_dual".into(), kv_cache_src);
461        sources.insert("kv_cache_copy_seq_f32_to_f16_kv_dual".into(), kv_cache_src);
462        // ADR-028 iter-145 — fused single-position K+V copy variants (decode shape)
463        sources.insert("kv_cache_copy_batch_f32_kv_dual".into(), kv_cache_src);
464        sources.insert("kv_cache_copy_batch_f32_to_f16_kv_dual".into(), kv_cache_src);
465        // bf16-source KV cache copy (Phase 2 bf16 activation path)
466        sources.insert("kv_cache_copy_seq_bf16".into(), kv_cache_src);
467        // ADR-030 iter-95 — bit-exact BF16→BF16 head-major cache copy for
468        // Option A xlen verify (avoids F16 round-trip precision drift).
469        sources.insert("kv_cache_copy_seq_bf16_to_bf16_head_major".into(), kv_cache_src);
470
471        // Elementwise and transpose kernels (Story 1.5)
472        let elementwise_src: &'static str = include_str!("shaders/elementwise.metal");
473        sources.insert("elementwise_add_f32".into(), elementwise_src);
474        sources.insert("elementwise_add_f16".into(), elementwise_src);
475        sources.insert("elementwise_mul_f32".into(), elementwise_src);
476        sources.insert("elementwise_mul_f16".into(), elementwise_src);
477        sources.insert("elementwise_add_bf16".into(), elementwise_src);
478        sources.insert("elementwise_mul_bf16".into(), elementwise_src);
479        sources.insert("cast_f16_to_f32".into(), elementwise_src);
480        sources.insert("cast_f32_to_f16".into(), elementwise_src);
481        sources.insert("cast_bf16_to_f32".into(), elementwise_src);
482        sources.insert("cast_f32_to_bf16".into(), elementwise_src);
483        sources.insert("scalar_mul_bf16".into(), elementwise_src);
484        sources.insert("scalar_mul_f32".into(), elementwise_src);
485        sources.insert("embedding_gather_scale_f32".into(), elementwise_src);
486        sources.insert("embedding_gather_scale_batch_f32".into(), elementwise_src);
487        sources.insert("permute_021_bf16".into(), elementwise_src);
488        sources.insert("transpose_last2_bf16".into(), elementwise_src);
489        sources.insert("transpose_last2_f16".into(), elementwise_src);
490        sources.insert("permute_021_f32".into(), elementwise_src);
491        sources.insert("permute_021_bf16_to_f32".into(), elementwise_src);
492        sources.insert("transpose_2d_f32".into(), elementwise_src);
493        sources.insert("transpose_2d_f16".into(), elementwise_src);
494
495        // Attention kernels (Story 1.3)
496        let sdpa_src: &'static str = include_str!("shaders/sdpa.metal");
497        sources.insert("sdpa".into(), sdpa_src);
498        sources.insert("sdpa_bf16".into(), sdpa_src);
499        let sdpa_sliding_src: &'static str = include_str!("shaders/sdpa_sliding.metal");
500        sources.insert("sdpa_sliding".into(), sdpa_sliding_src);
501        sources.insert("sdpa_sliding_bf16".into(), sdpa_sliding_src);
502
503        // Flash-attention tiled prefill kernel (ADR-011 Phase 1).
504        // Ten entry points; all backed by the same shader source.
505        // Pipelines are compiled with function constants via
506        // `get_pipeline_with_bool_constants` — not `get_pipeline`.
507        let flash_attn_prefill_src: &'static str =
508            include_str!("shaders/flash_attn_prefill.metal");
509        // D=256 variants (BQ=32, BK=16, WM=4, WN=1 — 128 threads/threadgroup)
510        sources.insert(
511            "steel_attention_float32_bq32_bk16_bd256_wm4_wn1_maskfloat32".into(),
512            flash_attn_prefill_src,
513        );
514        sources.insert(
515            "steel_attention_float32_bq32_bk16_bd256_wm4_wn1_maskbool_".into(),
516            flash_attn_prefill_src,
517        );
518        sources.insert(
519            "steel_attention_bfloat16_bq32_bk16_bd256_wm4_wn1_maskbfloat16".into(),
520            flash_attn_prefill_src,
521        );
522        sources.insert(
523            "steel_attention_bfloat16_bq32_bk16_bd256_wm4_wn1_maskbool_".into(),
524            flash_attn_prefill_src,
525        );
526        sources.insert(
527            "steel_attention_float16_bq32_bk16_bd256_wm4_wn1_maskfloat16".into(),
528            flash_attn_prefill_src,
529        );
530        sources.insert(
531            "steel_attention_float16_bq32_bk16_bd256_wm4_wn1_maskbool_".into(),
532            flash_attn_prefill_src,
533        );
534        // D=512 variants (BQ=8, BK=8, WM=1, WN=1 — 32 threads/threadgroup)
535        // NOTE: f32 at D=512 is NOT instantiated — threadgroup memory exceeds
536        // the 32 KB Metal limit (candle sdpa.rs:86-94).
537        sources.insert(
538            "steel_attention_bfloat16_bq8_bk8_bd512_wm1_wn1_maskbfloat16".into(),
539            flash_attn_prefill_src,
540        );
541        sources.insert(
542            "steel_attention_bfloat16_bq8_bk8_bd512_wm1_wn1_maskbool_".into(),
543            flash_attn_prefill_src,
544        );
545        sources.insert(
546            "steel_attention_float16_bq8_bk8_bd512_wm1_wn1_maskfloat16".into(),
547            flash_attn_prefill_src,
548        );
549        sources.insert(
550            "steel_attention_float16_bq8_bk8_bd512_wm1_wn1_maskbool_".into(),
551            flash_attn_prefill_src,
552        );
553
554        // Flash attention vector kernels — SIMD-vectorized decode-path SDPA
555        // (ported from llama.cpp flash_attn_ext_vec)
556        let flash_attn_vec_src: &'static str =
557            include_str!("shaders/flash_attn_vec.metal");
558        sources.insert("flash_attn_vec_dk256".into(), flash_attn_vec_src);
559        sources.insert("flash_attn_vec_dk512".into(), flash_attn_vec_src);
560        sources.insert("flash_attn_vec_reduce_dk256".into(), flash_attn_vec_src);
561        sources.insert("flash_attn_vec_reduce_dk512".into(), flash_attn_vec_src);
562        // F16 KV variants (Phase 4a)
563        sources.insert("flash_attn_vec_f16kv_dk256".into(), flash_attn_vec_src);
564        sources.insert("flash_attn_vec_f16kv_dk512".into(), flash_attn_vec_src);
565
566        // RoPE, normalization, activation kernels (Story 1.4)
567        let rope_src: &'static str = include_str!("shaders/rope.metal");
568        sources.insert("rope_f32".into(), rope_src);
569        sources.insert("rope_f16".into(), rope_src);
570        sources.insert("rope_bf16".into(), rope_src);
571        sources.insert("rope_neox_bf16".into(), rope_src);
572        sources.insert("rope_neox_f32".into(), rope_src);
573        let rms_norm_src: &'static str = include_str!("shaders/rms_norm.metal");
574        sources.insert("rms_norm_f32".into(), rms_norm_src);
575        // ADR-028 iter-310 — float4 + simd_sum variants (peer-pattern,
576        // ported from llama.cpp kernel_rms_norm_fuse_impl<float4, 1>).
577        // Env-gated via HF2Q_RMS_NORM_V2=1 in the dispatchers.
578        sources.insert("rms_norm_f32_v2".into(), rms_norm_src);
579        sources.insert("rms_norm_no_scale_f32_v2".into(), rms_norm_src);
580        sources.insert("rms_norm_f16".into(), rms_norm_src);
581        sources.insert("rms_norm_bf16".into(), rms_norm_src);
582        sources.insert("rms_norm_no_scale_bf16".into(), rms_norm_src);
583        sources.insert("rms_norm_no_scale_f32".into(), rms_norm_src);
584        sources.insert("rms_norm_no_scale_f32_dual".into(), rms_norm_src);
585        sources.insert("rms_norm_f32_triple".into(), rms_norm_src);
586        sources.insert("fused_post_attn_triple_norm_f32".into(), rms_norm_src);
587        // ADR-028 iter-370: V2 (float4 + simd_sum) variant of triple_norm.
588        sources.insert("fused_post_attn_triple_norm_f32_v2".into(), rms_norm_src);
589        // ADR-028 iter-217: fused post-FF norm 2 + end-of-layer FINAL
590        // (combines 2 sequential fused_norm_add dispatches into 1 kernel).
591        sources.insert("fused_post_ff_norm2_endlayer_f32".into(), rms_norm_src);
592        // ADR-028 iter-362: V2 (float4 + simd_sum) variant of the above.
593        // Same math, 75% fewer barriers per dispatch (4 vs 16 at tg=256).
594        sources.insert("fused_post_ff_norm2_endlayer_f32_v2".into(), rms_norm_src);
595        // ADR-028 iter-367: V2 fusion of moe_weighted_sum INTO Path A end-of-layer.
596        // Eliminates 1 dispatch + moe_accum round-trip from gemma4 decode default.
597        sources.insert("fused_moe_wsum_post_ff_norm2_endlayer_f32_v2".into(), rms_norm_src);
598        sources.insert("rms_norm_no_scale_f32_dual_perm".into(), rms_norm_src);
599        // Fused RMS norm + elementwise multiply kernels (Phase 4e.2)
600        sources.insert("rms_norm_mul_f32".into(), rms_norm_src);
601        sources.insert("rms_norm_mul_f16".into(), rms_norm_src);
602        sources.insert("rms_norm_mul_bf16".into(), rms_norm_src);
603        // L2 norm kernels (ADR-013 Decision 3 — Gated DeltaNet Q/K norm)
604        let l2_norm_src: &'static str = include_str!("shaders/l2_norm.metal");
605        sources.insert("l2_norm_f32".into(), l2_norm_src);
606        sources.insert("l2_norm_f16".into(), l2_norm_src);
607        sources.insert("l2_norm_bf16".into(), l2_norm_src);
608        // ADR-015 iter59a — fused L2 norm + scalar multiply (DN q-path).
609        sources.insert("l2_norm_scale_f32".into(), l2_norm_src);
610        // Cumulative-sum kernels (ADR-013 Decision 4 — DeltaNet decay-mask base)
611        let cumsum_src: &'static str = include_str!("shaders/cumsum.metal");
612        sources.insert("cumsum_f32".into(), cumsum_src);
613        sources.insert("cumsum_bf16".into(), cumsum_src);
614        // SSM conv kernels (ADR-013 Decision 7 — DeltaNet 1D causal conv + SiLU)
615        let ssm_conv_src: &'static str = include_str!("shaders/ssm_conv.metal");
616        sources.insert("ssm_conv_forward_f32".into(), ssm_conv_src);
617        sources.insert("ssm_conv_forward_bf16".into(), ssm_conv_src);
618        sources.insert("ssm_conv_state_update_f32".into(), ssm_conv_src);
619        sources.insert("ssm_conv_state_update_bf16".into(), ssm_conv_src);
620        // Tri-solve kernels (ADR-013 Decision 5 — chunked DeltaNet debug path)
621        let tri_solve_src: &'static str = include_str!("shaders/tri_solve.metal");
622        sources.insert("tri_solve_lower_unit_f32".into(), tri_solve_src);
623        sources.insert("tri_solve_lower_unit_bf16".into(), tri_solve_src);
624        // Rope-multi kernels (ADR-013 Decision 10 — IMROPE for Qwen3.5)
625        let rope_multi_src: &'static str = include_str!("shaders/rope_multi.metal");
626        sources.insert("rope_multi_f32".into(), rope_multi_src);
627        sources.insert("rope_multi_bf16".into(), rope_multi_src);
628        // Gated DeltaNet fused kernel (ADR-013 Decision 6 — centerpiece)
629        let gdn_src: &'static str = include_str!("shaders/gated_delta_net.metal");
630        sources.insert("gated_delta_net_f32".into(), gdn_src);
631        // ADR-015 iter56 — decode-only `simd_sum` variant. Three NSG-templated
632        // host names share the same source; selection is by D_k via
633        // `dispatch_gated_delta_net_decode`. Drop-in for the fused kernel
634        // above when n_tokens=1.
635        let gdn_decode_src: &'static str =
636            include_str!("shaders/gated_delta_net_decode.metal");
637        sources.insert("gated_delta_net_decode_f32_1".into(), gdn_decode_src);
638        sources.insert("gated_delta_net_decode_f32_2".into(), gdn_decode_src);
639        sources.insert("gated_delta_net_decode_f32_4".into(), gdn_decode_src);
640        // Wave 5b — chunk-parallel inter-chunk state-recurrence kernel
641        // (the one new kernel in the chunk-parallel pipeline; spec source:
642        // arXiv 2412.06464 §4 + FLA chunk_delta_h.py:43-298).
643        let gdn_chunk_src: &'static str =
644            include_str!("shaders/gated_delta_net_chunk.metal");
645        sources.insert(
646            "gated_delta_net_chunk_inter_state_bf16".into(),
647            gdn_chunk_src,
648        );
649        // Wave 5b.1 iter 2 — chunk_scaled_dot_kkt kernel (input-side of
650        // the chunk pipeline; spec source: FLA chunk_scaled_dot_kkt.py:36-99).
651        let gdn_kkt_src: &'static str =
652            include_str!("shaders/gated_delta_net_kkt.metal");
653        sources.insert("gated_delta_net_kkt_bf16".into(), gdn_kkt_src);
654        // Wave 5b.1 iter 2 — recompute_w_u_fwd kernel (applies post-solve A
655        // to (β·v) and (β·k·exp(g)) to produce w and u; spec source: FLA
656        // wy_fast.py:29-117).
657        let gdn_recompute_wu_src: &'static str =
658            include_str!("shaders/gated_delta_net_recompute_wu.metal");
659        sources.insert(
660            "gated_delta_net_recompute_wu_bf16".into(),
661            gdn_recompute_wu_src,
662        );
663        // Wave 5b.1 iter 3 — chunk_fwd_o kernel (per-chunk output: closes
664        // the chunk pipeline; spec source: FLA chunk_o.py:42-138).
665        let gdn_chunk_o_src: &'static str =
666            include_str!("shaders/gated_delta_net_chunk_o.metal");
667        sources.insert("gated_delta_net_chunk_o_bf16".into(), gdn_chunk_o_src);
668        // Wave 5b.1 iter 4 — orchestrator helper kernels:
669        //   chunk_local_cumsum_g_f32      — per-chunk prefix sum on g [B, T, H]
670        //   chunk_tri_solve_invert_f32    — per-chunk-block (I + A_strict)^-1
671        //                                   on FLA's [B, T, H, BT] layout.
672        let chunk_local_cumsum_g_src: &'static str =
673            include_str!("shaders/chunk_local_cumsum_g.metal");
674        sources.insert(
675            "chunk_local_cumsum_g_f32".into(),
676            chunk_local_cumsum_g_src,
677        );
678        let chunk_tri_solve_invert_src: &'static str =
679            include_str!("shaders/chunk_gated_delta_rule_tri_solve_invert.metal");
680        sources.insert(
681            "chunk_tri_solve_invert_f32".into(),
682            chunk_tri_solve_invert_src,
683        );
684        // Sigmoid-gated elementwise multiply (ADR-013 Decision 9 — full-attn output gate)
685        let sigmoid_mul_src: &'static str = include_str!("shaders/sigmoid_mul.metal");
686        sources.insert("sigmoid_mul_f32".into(), sigmoid_mul_src);
687        sources.insert("sigmoid_mul_bf16".into(), sigmoid_mul_src);
688        let silu_mul_src: &'static str = include_str!("shaders/silu_mul.metal");
689        sources.insert("silu_mul_f32".into(), silu_mul_src);
690        let compute_g_beta_src: &'static str = include_str!("shaders/compute_g_beta.metal");
691        sources.insert("compute_g_beta_f32".into(), compute_g_beta_src);
692        let ssm_norm_gate_src: &'static str = include_str!("shaders/ssm_norm_gate.metal");
693        sources.insert("ssm_norm_gate_f32".into(), ssm_norm_gate_src);
694        let gelu_src: &'static str = include_str!("shaders/gelu.metal");
695        sources.insert("gelu_f32".into(), gelu_src);
696        sources.insert("gelu_f16".into(), gelu_src);
697        sources.insert("gelu_bf16".into(), gelu_src);
698        let softmax_src: &'static str = include_str!("shaders/softmax.metal");
699        sources.insert("softmax_f32".into(), softmax_src);
700        sources.insert("softmax_f16".into(), softmax_src);
701        sources.insert("softmax_bf16".into(), softmax_src);
702        let softmax_backward_src: &'static str =
703            include_str!("shaders/softmax_backward.metal");
704        sources.insert("softmax_backward_f32".into(), softmax_backward_src);
705        let log_elementwise_src: &'static str =
706            include_str!("shaders/log_elementwise.metal");
707        sources.insert("log_f32".into(), log_elementwise_src);
708        sources.insert("log_backward_f32".into(), log_elementwise_src);
709        let row_sum_src: &'static str = include_str!("shaders/row_sum.metal");
710        sources.insert("row_sum_f32".into(), row_sum_src);
711        sources.insert("row_sum_backward_f32".into(), row_sum_src);
712        // ADR-020 iter-10a: GGUF-legacy quantize-dequantize round-trip kernels
713        // (Q4_0 + Q8_0).  Used by hf2q's dynamic_quant Track 1 to produce
714        // W_low / W_high for the gradient-Taylor sensitivity formula.
715        let qdq_legacy_src: &'static str = include_str!("shaders/qdq_legacy.metal");
716        sources.insert("qdq_q4_0_f32".into(), qdq_legacy_src);
717        sources.insert("qdq_q8_0_f32".into(), qdq_legacy_src);
718        // ADR-020 iter-10b: RMSNorm reverse-mode autograd kernels.
719        // r_inv helper is reused by both backward kernels; dx and dw cover
720        // the full backward identity for `y = x * rsqrt(mean(x²) + eps) * w`.
721        let rms_norm_backward_src: &'static str =
722            include_str!("shaders/rms_norm_backward.metal");
723        sources.insert(
724            "rms_norm_compute_rms_inv_f32".into(),
725            rms_norm_backward_src,
726        );
727        sources.insert("rms_norm_backward_dx_f32".into(), rms_norm_backward_src);
728        sources.insert("rms_norm_backward_dw_f32".into(), rms_norm_backward_src);
729        // ADR-020 iter-11a: 2-D row-major slice + concat-by-column kernels.
730        // Used by hf2q's multi-head SDPA on GpuTape (slice Q/K/V into
731        // per-head views, run per-head SDPA, concat per-head contexts
732        // back to full attention output).
733        let slice_concat_2d_src: &'static str =
734            include_str!("shaders/slice_concat_2d.metal");
735        sources.insert("slice_2d_cols_f32".into(), slice_concat_2d_src);
736        sources.insert("copy_2d_cols_into_f32".into(), slice_concat_2d_src);
737        // ADR-020 iter-11b: SiLU forward + backward kernels for GpuTape
738        // SwiGLU FFN composition.
739        let silu_backward_src: &'static str =
740            include_str!("shaders/silu_backward.metal");
741        sources.insert("silu_f32".into(), silu_backward_src);
742        sources.insert("silu_backward_f32".into(), silu_backward_src);
743        // ADR-020 iter-11d: FP32 embedding lookup + scatter-add backward.
744        let embedding_autograd_src: &'static str =
745            include_str!("shaders/embedding_autograd.metal");
746        sources.insert("embedding_lookup_f32".into(), embedding_autograd_src);
747        sources.insert(
748            "embedding_scatter_add_f32".into(),
749            embedding_autograd_src,
750        );
751        // ADR-020 iter-13a: Adam optimizer step kernel for Track 2
752        // DWQ-proper training loop.
753        let adam_update_src: &'static str =
754            include_str!("shaders/adam_update.metal");
755        sources.insert("adam_update_f32".into(), adam_update_src);
756        // ADR-020 iter-13b: differentiable affine qdq kernels for the
757        // DWQ-proper training loop.  Init + forward + backward (scales,
758        // biases) — q_int is FROZEN, scales+biases learnable.
759        let qdq_affine_src: &'static str =
760            include_str!("shaders/qdq_affine.metal");
761        sources.insert("qdq_affine_init_f32".into(), qdq_affine_src);
762        sources.insert("qdq_affine_forward_f32".into(), qdq_affine_src);
763        sources.insert(
764            "qdq_affine_backward_scales_f32".into(),
765            qdq_affine_src,
766        );
767        sources.insert(
768            "qdq_affine_backward_biases_f32".into(),
769            qdq_affine_src,
770        );
771        // ADR-020 iter-15: fused affine quantized matmul for DWQ inference.
772        // Per-element kernel; one thread per (m, n) output element.
773        // Tiled + simdgroup-MMA variant lands in iter-15b.
774        let qmm_affine_src: &'static str =
775            include_str!("shaders/qmm_affine.metal");
776        sources.insert("qmm_affine_t_f32".into(), qmm_affine_src);
777        // ADR-020 iter-15b: tiled variant — 16x16 thread block with
778        // cooperative-load X/W tiles in threadgroup-shared memory for
779        // 2-5x speedup over the per-element kernel.
780        let qmm_affine_tiled_src: &'static str =
781            include_str!("shaders/qmm_affine_tiled.metal");
782        sources.insert(
783            "qmm_affine_t_f32_tiled".into(),
784            qmm_affine_tiled_src,
785        );
786        // ADR-020 iter-15c: simdgroup-MMA variant — uses Apple GPU
787        // hardware `simdgroup_matrix<float, 8, 8>` MMA for the inner
788        // reduction.  Per-tile algorithmic 8× over scalar tiled, lands
789        // as ~3-4× wall after launch / load amortization.
790        let qmm_affine_simd_src: &'static str =
791            include_str!("shaders/qmm_affine_simd.metal");
792        sources.insert(
793            "qmm_affine_t_f32_simd".into(),
794            qmm_affine_simd_src,
795        );
796        // ADR-020 iter-15c-2: 4-simdgroup-per-TG variant — 32×32
797        // output tile, 4 simdgroups arranged as 2×2 grid each owning
798        // a 16×16 sub-tile = 4 simdgroup_matrix accumulators.  Same
799        // math as 15c-1, fuller warp-pool exploitation.
800        let qmm_affine_simd4_src: &'static str =
801            include_str!("shaders/qmm_affine_simd4.metal");
802        sources.insert(
803            "qmm_affine_t_f32_simd4".into(),
804            qmm_affine_simd4_src,
805        );
806        // ADR-020 iter-15c-2b: gs=64 variant (mlx-lm dynamic_quant
807        // canonical default).  Same 4-simdgroup geometry, BK=64
808        // instead of 32 (= 8 sub-K-tiles per K-step instead of 4).
809        let qmm_affine_simd4_gs64_src: &'static str =
810            include_str!("shaders/qmm_affine_simd4_gs64.metal");
811        sources.insert(
812            "qmm_affine_t_f32_simd4_gs64".into(),
813            qmm_affine_simd4_gs64_src,
814        );
815        // ADR-020 AC#5 Iter A: packed-U32 dense affine matmul (bits=4,
816        // gs=32) — production decode/prefill kernel for serving DWQ
817        // safetensors directly without a load-time unpack pass.
818        let qmm_affine_t_packed_simd4_b4_src: &'static str =
819            include_str!("shaders/qmm_affine_t_packed_simd4_b4.metal");
820        sources.insert(
821            "qmm_affine_t_packed_simd4_b4".into(),
822            qmm_affine_t_packed_simd4_b4_src,
823        );
824        // ADR-020 iter-11h-b: training-mode causal depthwise 1D
825        // convolution (forward + backward dx + backward dw).  Used by
826        // GpuTape autograd for differentiable Qwen3.5MoE forward
827        // (GatedDeltaNet's conv1d step).
828        let conv1d_dwc_src: &'static str =
829            include_str!("shaders/conv1d_depthwise_causal.metal");
830        sources.insert(
831            "conv1d_depthwise_causal_forward_f32".into(),
832            conv1d_dwc_src,
833        );
834        sources.insert(
835            "conv1d_depthwise_causal_backward_dx_f32".into(),
836            conv1d_dwc_src,
837        );
838        sources.insert(
839            "conv1d_depthwise_causal_backward_dw_f32".into(),
840            conv1d_dwc_src,
841        );
842        // ADR-020 iter-11h-c1: elementwise exp forward + backward.
843        // Building block for GatedDeltaNet's alpha = exp(-g) state-decay.
844        let exp_src: &'static str =
845            include_str!("shaders/exp_elementwise.metal");
846        sources.insert("exp_f32".into(), exp_src);
847        sources.insert("exp_backward_f32".into(), exp_src);
848        // ADR-020 iter-11h-c2: vector outer product (forward + dlhs +
849        // drhs).  Building block for gated_delta_update's
850        // outer(delta, k) state-update term.
851        let outer_src: &'static str =
852            include_str!("shaders/outer_product.metal");
853        sources.insert("outer_product_f32".into(), outer_src);
854        sources.insert("outer_product_backward_lhs_f32".into(), outer_src);
855        sources.insert("outer_product_backward_rhs_f32".into(), outer_src);
856        // ADR-020 iter-11h-e1: take_along_axis (gather) + scatter-backward.
857        // Building block for MoE router on GpuTape.
858        let taa_src: &'static str =
859            include_str!("shaders/take_along_axis.metal");
860        sources.insert("take_along_axis_f32".into(), taa_src);
861        sources.insert("take_along_axis_backward_f32".into(), taa_src);
862        // ADR-020 iter-11h-misc-1: elementwise divide forward + backward.
863        let div_src: &'static str =
864            include_str!("shaders/divide_elementwise.metal");
865        sources.insert("divide_f32".into(), div_src);
866        sources.insert("divide_backward_f32".into(), div_src);
867        // ADR-020 iter-11h-misc-3: elementwise sqrt forward + backward.
868        let sqrt_src: &'static str =
869            include_str!("shaders/sqrt_elementwise.metal");
870        sources.insert("sqrt_f32".into(), sqrt_src);
871        sources.insert("sqrt_backward_f32".into(), sqrt_src);
872        let softcap_src: &'static str = include_str!("shaders/softcap.metal");
873        sources.insert("softcap_f32".into(), softcap_src);
874        sources.insert("softcap_f16".into(), softcap_src);
875        sources.insert("softcap_bf16".into(), softcap_src);
876
877        // Fused norm-add kernels — Gemma4 post-attention / post-FFN ordering:
878        //   normed = rms_norm(input, weight, eps);  output = residual + normed
879        let fused_norm_add_src: &'static str =
880            include_str!("shaders/fused_norm_add_bf16.metal");
881        sources.insert("fused_norm_add_bf16".into(), fused_norm_add_src);
882        sources.insert("fused_norm_add_no_weight_bf16".into(), fused_norm_add_src);
883
884        // Fused head-norm + RoPE f32 kernel — replaces separate rms_norm + rope_neox_f32
885        let fused_hnr_f32_src: &'static str =
886            include_str!("shaders/fused_head_norm_rope_f32.metal");
887        sources.insert("fused_head_norm_rope_f32".into(), fused_hnr_f32_src);
888        // ADR-028 iter-337 — float4 + simd_sum Phase 1 variant.  Phases
889        // 2-4 byte-identical to v1; race-fix barrier preserved.  Env-gated
890        // via HF2Q_FUSED_HEAD_NORM_ROPE_V2 (default ON, opt-out via =0).
891        sources.insert("fused_head_norm_rope_f32_v2".into(), fused_hnr_f32_src);
892
893        // Fused head-norm + RoPE bf16 kernels (single-token + batch prefill)
894        // Both entry points live in the same .metal file.
895        let fused_hnr_bf16_src: &'static str =
896            include_str!("shaders/fused_head_norm_rope_bf16.metal");
897        sources.insert("fused_head_norm_rope_bf16".into(), fused_hnr_bf16_src);
898        sources.insert("fused_head_norm_rope_batch_bf16".into(), fused_hnr_bf16_src);
899
900        // Fused norm-add f32 kernels — post-attention / post-FFN / end-of-layer
901        let fused_norm_add_f32_src: &'static str =
902            include_str!("shaders/fused_norm_add_f32.metal");
903        sources.insert("fused_norm_add_f32".into(), fused_norm_add_f32_src);
904        // ADR-028 iter-331 — float4 + simd_sum variant (peer-pattern,
905        // ported from llama.cpp kernel_rms_norm_fuse_impl<float4, 3>).
906        // Env-gated via HF2Q_FUSED_NORM_ADD_V2=1 in the dispatcher
907        // (default ON since iter-331; opt-out via =0/false/off).
908        sources.insert("fused_norm_add_f32_v2".into(), fused_norm_add_f32_src);
909        sources.insert("fused_residual_norm_f32".into(), fused_norm_add_f32_src);
910        sources.insert("fused_residual_norm_scalar_f32".into(), fused_norm_add_f32_src);
911        sources.insert("fused_moe_routing_f32".into(), fused_norm_add_f32_src);
912        // ADR-028 iter-363: V2 (simd_max + simd_sum) variant of MoE routing.
913        sources.insert("fused_moe_routing_f32_v2".into(), fused_norm_add_f32_src);
914        // ADR-029 iter-175 Step 1i: V3 = V2 + parallel SG-tournament top-K
915        // (replaces V2's single-thread serial scan for k = 0..top_k).
916        sources.insert("fused_moe_routing_f32_v3".into(), fused_norm_add_f32_src);
917        sources.insert("fused_moe_routing_batch_f32".into(), fused_norm_add_f32_src);
918        // ADR-029 iter-175 Step 1j: batched-prefill V3 (same parallel
919        // SG-tournament top-K as fused_moe_routing_f32_v3, applied per-token
920        // within each TG of the batched dispatch).
921        sources.insert("fused_moe_routing_batch_f32_v3".into(), fused_norm_add_f32_src);
922        sources.insert("fused_norm_add_scalar_f32".into(), fused_norm_add_f32_src);
923        sources.insert("fused_moe_wsum_norm_add_f32".into(), fused_norm_add_f32_src);
924        sources.insert("fused_moe_wsum_dnorm_add_f32".into(), fused_norm_add_f32_src);
925
926        // Argsort kernel (Story 2.3) — MoE top-K routing
927        let argsort_src: &'static str = include_str!("shaders/argsort.metal");
928        sources.insert("argsort_desc_f32".into(), argsort_src);
929
930        // Gather / index_select kernel (Story 2.4)
931        let gather_src: &'static str = include_str!("shaders/gather.metal");
932        sources.insert("gather_f32".into(), gather_src);
933
934        // F32 KV cache copy kernel (Session merge S1+S2)
935        let kv_cache_copy_src: &'static str =
936            include_str!("shaders/kv_cache_copy.metal");
937        sources.insert("kv_cache_copy".into(), kv_cache_copy_src);
938        sources.insert("kv_cache_copy_f32".into(), kv_cache_copy_src);
939
940        // Strided copy kernel (Story 2.5)
941        let copy_src: &'static str = include_str!("shaders/copy.metal");
942        sources.insert("strided_copy_f32".into(), copy_src);
943        sources.insert("offset_copy_f32".into(), copy_src);
944
945        // Fused-QKV split kernel (ADR-005 W-5b.18 — replaces hf2q CPU
946        // download → triple-loop split → 3× upload round-trip in
947        // gpu_delta_net::layer_qkv_deinterleave).
948        let qkv_split_src: &'static str = include_str!("shaders/qkv_split.metal");
949        sources.insert("qkv_split_f32".into(), qkv_split_src);
950
951        // Tiled-GQA broadcast kernel (ADR-005 W-5b.19 — replaces hf2q CPU
952        // tiled-replicate at gpu_delta_net::apply_gated_delta_net_chunk
953        // GQA pre-expansion, ~497 ms / 10.4 ms-per-layer at PP4106).
954        let repeat_tiled_src: &'static str =
955            include_str!("shaders/repeat_tiled.metal");
956        sources.insert("repeat_tiled_f32".into(), repeat_tiled_src);
957
958        // Dense F16 GEMM kernel (Story 2.6) — lm_head projection
959        let dense_gemm_src: &'static str = include_str!("shaders/dense_gemm.metal");
960        sources.insert("dense_gemm_f16".into(), dense_gemm_src);
961        sources.insert("dense_matvec_f16".into(), dense_gemm_src);
962        sources.insert("dense_matvec_f16w_f32io".into(), dense_gemm_src);
963        // BF16-weight mat-vec: BF16 weights × F32 input → F32 output (decode lm_head)
964        sources.insert("dense_matvec_bf16w_f32io".into(), dense_gemm_src);
965        // Pure F32 mat-vec: F32 weights × F32 input → F32 output (decode lm_head)
966        sources.insert("dense_matvec_f32".into(), dense_gemm_src);
967
968        // Standalone FWHT for TurboQuant pre/post-rotation (SIMD shuffle, zero barriers)
969        let fwht_src: &'static str = include_str!("shaders/fwht_standalone.metal");
970        sources.insert("fwht_standalone_f32_d256".into(), fwht_src);
971        sources.insert("fwht_standalone_f32_d512".into(), fwht_src);
972        // ADR-007 iter-14 D1 SRHT variants: sign pre-mult (for Q) + sign undo (for output)
973        sources.insert("fwht_sign_premult_f32_d256".into(), fwht_src);
974        sources.insert("fwht_sign_premult_f32_d512".into(), fwht_src);
975        sources.insert("fwht_sign_undo_f32_d256".into(), fwht_src);
976        sources.insert("fwht_sign_undo_f32_d512".into(), fwht_src);
977
978        // Fast Hadamard quantize (SIMD shuffle, zero barriers)
979        let hq_fast_src: &'static str = include_str!("shaders/hadamard_quantize_kv_fast.metal");
980        sources.insert("hadamard_quantize_kv_fast_d256".into(), hq_fast_src);
981        sources.insert("hadamard_quantize_kv_fast_d512".into(), hq_fast_src);
982        // ADR-028 iter-485 (Phase 7d / H4): fused K+V single-position 4-bit encoder.
983        sources.insert("hadamard_quantize_kv_fast_dual_d256".into(), hq_fast_src);
984        sources.insert("hadamard_quantize_kv_fast_dual_d512".into(), hq_fast_src);
985        // Track B (iter-21): higher-bit (5/6-bit) quantize kernels (byte-packed)
986        sources.insert("hadamard_quantize_kv_hb_d256".into(), hq_fast_src);
987        sources.insert("hadamard_quantize_kv_hb_d512".into(), hq_fast_src);
988        // ADR-028 iter-148: fused K+V single-position HB encoder
989        sources.insert("hadamard_quantize_kv_hb_dual_d256".into(), hq_fast_src);
990        sources.insert("hadamard_quantize_kv_hb_dual_d512".into(), hq_fast_src);
991        // ADR-028 Phase 10e.5 (iter-351): no-FWHT V quantize for hybrid path.
992        // Same byte-packed Lloyd-Max codebook output, but skips the Hadamard
993        // rotation so dequant in SDPA recovers raw V (no FWHT-undo needed).
994        sources.insert("kv_quantize_v_no_fwht_d256".into(), hq_fast_src);
995        sources.insert("kv_quantize_v_no_fwht_d512".into(), hq_fast_src);
996        // ADR-028 Phase 10c.5 (iter-354): fused F16-K-copy + V-no-FWHT-encode.
997        // Saves 30 KV-write dispatches/decode-token at gemma4 30L by combining
998        // the per-layer K-cast and V-encode into a single dispatch (Z-dim).
999        sources.insert("kv_copy_kf16_quantize_v_no_fwht_d256".into(), hq_fast_src);
1000        sources.insert("kv_copy_kf16_quantize_v_no_fwht_d512".into(), hq_fast_src);
1001
1002        // iter-20 Leg F: TQ KV dequantize kernel (nibbles+norms → F32)
1003        let tq_dq_src: &'static str = include_str!("shaders/tq_dequantize_kv.metal");
1004        sources.insert("tq_dequantize_kv".into(), tq_dq_src);
1005        // Track B (iter-21): higher-bit dequantize kernel (byte-packed indices)
1006        sources.insert("tq_dequantize_hb_kv".into(), tq_dq_src);
1007        // ADR-027 Phase B iter-30 (hf2q sub-sub-iter 23c-β.1): sequence-batch
1008        // dequant variant. Same MSL source; new kernel entry point
1009        // `tq_dequantize_hb_kv_seq` reads positions [start_pos..start_pos+n_tokens)
1010        // in one dispatch (one threadgroup per (kv_head, position)). Unblocks
1011        // hf2q's TQ-aware prefill SDPA path (current per-position kernel
1012        // requires cur_len separate dispatches).
1013        sources.insert("tq_dequantize_hb_kv_seq".into(), tq_dq_src);
1014
1015        // iter-24: native higher-bit (5/6/8-bit) TQ SDPA kernel (byte-packed K/V)
1016        let tq_hb_src: &'static str = include_str!("shaders/flash_attn_vec_tq_hb.metal");
1017        sources.insert("flash_attn_vec_tq_hb_dk256".into(), tq_hb_src);
1018        sources.insert("flash_attn_vec_tq_hb_dk512".into(), tq_hb_src);
1019
1020        // ADR-028 §iter-485 (Phase 7d H3): fused TQ-HB reduce + FWHT-sign-undo.
1021        // Combines flash_attn_vec_reduce + fwht_sign_undo_f32 into a single
1022        // dispatch, saving 1 dispatch + 1 forced barrier per layer per decode
1023        // token. Gated by env flag `HF2Q_TQ_HB_OUT_FUSED=1` in forward_mlx.rs.
1024        let reduce_undo_src: &'static str = include_str!("shaders/flash_attn_vec_reduce_tq_hb_undo.metal");
1025        sources.insert("flash_attn_vec_reduce_tq_hb_undo_dk256".into(), reduce_undo_src);
1026        sources.insert("flash_attn_vec_reduce_tq_hb_undo_dk512".into(), reduce_undo_src);
1027
1028        // ADR-028 Phase 10d (iter-349): hybrid F16-K + TQ-HB-V SDPA kernel.
1029        // Same V-side codebook as flash_attn_vec_tq_hb (5/6/8-bit Lloyd-Max);
1030        // K-side reads F16 dense — peer-equivalent layout, no codebook lookup.
1031        let hybrid_src: &'static str = include_str!("shaders/flash_attn_vec_hybrid.metal");
1032        sources.insert("flash_attn_vec_hybrid_dk256".into(), hybrid_src);
1033        sources.insert("flash_attn_vec_hybrid_dk512".into(), hybrid_src);
1034
1035        // ADR-029 CFA cfa-20260512-fa-peer-port (iter-122): verbatim llama.cpp peer port.
1036        // F16-K + F16-V, DK=DV=256, NWG=1, NSG=1, NE=1. No function constants — baked.
1037        let peer_port_src: &'static str = include_str!("shaders/flash_attn_vec_peer_port_f16.metal");
1038        sources.insert("flash_attn_vec_peer_port_f16_dk256_dv256".into(), peer_port_src);
1039
1040        // ADR-029 iter-134: peer reduce kernel (verbatim port of ggml-metal.metal 7235-7275).
1041        // Pairs with the NWG=32 vec kernel to match peer's actual runtime dispatch.
1042        let peer_port_reduce_src: &'static str =
1043            include_str!("shaders/flash_attn_vec_peer_port_f16_reduce.metal");
1044        sources.insert(
1045            "flash_attn_vec_peer_port_f16_reduce_dv256_nwg32".into(),
1046            peer_port_reduce_src,
1047        );
1048
1049        // ADR-029 iter-135: NWG=32 variant of the verbatim peer port. Same body as
1050        // flash_attn_vec_peer_port_f16.metal with NWG=1→32. Pairs with iter-134 reduce kernel.
1051        let peer_port_nwg32_src: &'static str =
1052            include_str!("shaders/flash_attn_vec_peer_port_f16_nwg32.metal");
1053        sources.insert(
1054            "flash_attn_vec_peer_port_f16_nwg32_dk256_dv256".into(),
1055            peer_port_nwg32_src,
1056        );
1057
1058        // GPU sampling kernels — eliminate logits readback (Phase 6)
1059        let argmax_src: &'static str = include_str!("shaders/argmax.metal");
1060        sources.insert("argmax_f32".into(), argmax_src);
1061        let softmax_sample_src: &'static str =
1062            include_str!("shaders/softmax_sample.metal");
1063        sources.insert("softmax_sample_f32".into(), softmax_sample_src);
1064        // Top-K kernel for Q8 rerank: avoids full-logits readback.
1065        let top_k_src: &'static str = include_str!("shaders/top_k.metal");
1066        sources.insert("top_k_f32".into(), top_k_src);
1067
1068        // MoE GPU routing + weighted reduce (ADR-013 P13.3 perf).
1069        // Replaces CPU softmax+topk round-trip and CPU weighted accumulate.
1070        let moe_stk_src: &'static str =
1071            include_str!("shaders/moe_softmax_topk.metal");
1072        sources.insert("moe_softmax_topk_f32".into(), moe_stk_src);
1073        let moe_wr_src: &'static str =
1074            include_str!("shaders/moe_weighted_reduce.metal");
1075        sources.insert("moe_weighted_reduce_f32".into(), moe_wr_src);
1076        let sdpa_decode_src: &'static str =
1077            include_str!("shaders/sdpa_decode.metal");
1078        sources.insert("sdpa_decode".into(), sdpa_decode_src);
1079
1080        Self {
1081            cache: HashMap::new(),
1082            sources,
1083            precompiled_lib: None,
1084            precompiled_load_attempted: false,
1085        }
1086    }
1087
1088    /// Try to obtain the precompiled `default.metallib` Library, loading it
1089    /// lazily on first call.  Returns `None` when:
1090    /// - `MLX_PRECOMPILED_METALLIB` is unset (default)
1091    /// - The embedded blob is empty (build.rs skipped metallib build)
1092    /// - `device.new_library_with_data` failed previously
1093    /// - The previous load attempt already failed (no retry)
1094    fn try_precompiled_lib(
1095        &mut self,
1096        device: &metal::DeviceRef,
1097    ) -> Option<&metal::LibraryRef> {
1098        if !precompiled_enabled() {
1099            return None;
1100        }
1101        if !self.precompiled_load_attempted {
1102            self.precompiled_load_attempted = true;
1103            if EMBEDDED_METALLIB.is_empty() {
1104                return None;
1105            }
1106            // Apple's `newLibraryWithData:` expects a dispatch_data_t.
1107            // metal-rs wraps this via `new_library_with_data` which takes
1108            // a `&[u8]`.
1109            match device.new_library_with_data(EMBEDDED_METALLIB) {
1110                Ok(lib) => self.precompiled_lib = Some(lib),
1111                Err(_) => self.precompiled_lib = None,
1112            }
1113        }
1114        self.precompiled_lib.as_deref()
1115    }
1116
1117    /// Register a shader source at runtime (useful for testing and dynamic
1118    /// kernel generation).
1119    pub fn register_source(&mut self, name: impl Into<String>, source: &'static str) {
1120        let name = name.into();
1121        // Invalidate any cached pipeline for this name since the source changed.
1122        self.cache.remove(&name);
1123        self.sources.insert(name, source);
1124    }
1125
1126    /// Get a compiled compute pipeline for the named kernel function.
1127    ///
1128    /// On first call for a given name, this compiles the MSL source into a
1129    /// Metal library, extracts the named function, and creates a
1130    /// `ComputePipelineState`.  Subsequent calls return the cached pipeline.
1131    ///
1132    /// # Errors
1133    ///
1134    /// * `MlxError::KernelNotFound` — no source registered for this name.
1135    /// * `MlxError::ShaderCompilationError` — MSL compilation or pipeline
1136    ///   creation failed.
1137    pub fn get_pipeline(
1138        &mut self,
1139        name: &str,
1140        device: &metal::DeviceRef,
1141    ) -> Result<&ComputePipelineState> {
1142        if !self.cache.contains_key(name) {
1143            // ADR-029 iter-175 Step 1l: precompiled .metallib fast path.
1144            // When MLX_PRECOMPILED_METALLIB=1 AND the kernel exists in the
1145            // embedded library, use it.  Otherwise fall through to runtime
1146            // source compile.  Empirically ~+5.89% faster on q6_K matvec
1147            // (iter 1k bench).
1148            let precompiled_function = self
1149                .try_precompiled_lib(device)
1150                .and_then(|lib| lib.get_function(name, None).ok());
1151
1152            let function = match precompiled_function {
1153                Some(f) => f,
1154                None => {
1155                    // Slow path: compile the shader.
1156                    let source = self.sources.get(name).ok_or_else(|| {
1157                        MlxError::KernelNotFound(name.to_string())
1158                    })?;
1159
1160                    let compile_opts = metal::CompileOptions::new();
1161                    let library = device
1162                        .new_library_with_source(source, &compile_opts)
1163                        .map_err(|msg| MlxError::ShaderCompilationError {
1164                            name: name.to_string(),
1165                            message: msg,
1166                        })?;
1167
1168                    library
1169                        .get_function(name, None)
1170                        .map_err(|msg| MlxError::ShaderCompilationError {
1171                            name: name.to_string(),
1172                            message: msg,
1173                        })?
1174                }
1175            };
1176
1177            // Build the pipeline through a descriptor so we can attach a
1178            // human-readable label.  The label propagates into Instruments /
1179            // xctrace Metal System Trace as the per-pipeline identifier
1180            // (`metal-object-label` schema), giving us per-kernel attribution
1181            // instead of the generic "Compute Command 0" placeholder.
1182            //
1183            // `MTLComputePipelineState.label` is read-only after creation per
1184            // the Apple Metal spec; the only supported way to set it is via
1185            // the descriptor before pipeline creation.  ADR-015 iter9b.
1186            let descriptor = ComputePipelineDescriptor::new();
1187            descriptor.set_compute_function(Some(&function));
1188            descriptor.set_label(name);
1189            // ADR-028 iter-376: threadGroupSizeIsMultipleOfThreadExecutionWidth
1190            // hint allows the Metal compiler to skip bounds checks and use more
1191            // aggressive codegen.  Opt-in via HF2Q_PIPELINE_TG_MULT_HINT=1.
1192            // SAFETY: every dispatched threadgroup MUST be a multiple of 32 at
1193            // runtime — Apple specifies undefined behavior otherwise.  Our hot
1194            // kernels use tg_size ∈ {32, 64, 256, 1024} (all multiples of 32).
1195            if std::env::var("HF2Q_PIPELINE_TG_MULT_HINT").ok().as_deref() == Some("1") {
1196                descriptor.set_thread_group_size_is_multiple_of_thread_execution_width(true);
1197            }
1198
1199            let pipeline = device
1200                .new_compute_pipeline_state(&descriptor)
1201                .map_err(|msg| MlxError::ShaderCompilationError {
1202                    name: name.to_string(),
1203                    message: msg,
1204                })?;
1205
1206            self.cache.insert(name.to_string(), pipeline);
1207        }
1208
1209        // At this point the pipeline is guaranteed to be in the cache.
1210        // We use `ok_or_else` instead of `expect` to satisfy the no-panic policy.
1211        self.cache.get(name).ok_or_else(|| {
1212            MlxError::KernelNotFound(name.to_string())
1213        })
1214    }
1215
1216    /// Get a compiled compute pipeline for the named kernel, specialized with
1217    /// Metal function constants (both bool and i32 in one call).
1218    ///
1219    /// `bool_constants` contains `(index, value)` pairs mapping to
1220    /// `[[function_constant(index)]]` bool declarations in the MSL shader.
1221    /// `int_constants` contains `(index, value)` pairs mapping to
1222    /// `[[function_constant(index)]]` int (int32_t) declarations in the MSL
1223    /// shader.
1224    ///
1225    /// Pipelines are cached by a composite key:
1226    /// `"<name>|<index>:b<0|1>|...|<index>:i<value>|..."`.  The 'b' prefix
1227    /// marks bool entries and the 'i' prefix marks i32 entries, making the
1228    /// format unambiguous regardless of constant ordering.  Distinct
1229    /// `(name, constants)` combinations each compile to a separate pipeline;
1230    /// the slow compilation path runs at most once per unique combination.
1231    ///
1232    /// # Errors
1233    ///
1234    /// * `MlxError::KernelNotFound` — no source registered for this name.
1235    /// * `MlxError::ShaderCompilationError` — MSL compilation, function
1236    ///   specialisation, or pipeline creation failed.
1237    pub fn get_pipeline_with_constants(
1238        &mut self,
1239        name: &str,
1240        device: &metal::DeviceRef,
1241        bool_constants: &[(usize, bool)],
1242        int_constants: &[(usize, i32)],
1243    ) -> Result<&ComputePipelineState> {
1244        // Build a composite cache key so distinct constant combinations each
1245        // compile to their own pipeline.  Bool entries use the 'b' type marker
1246        // and i32 entries use 'i'; this prevents a collision between, e.g.,
1247        // bool index 5 value 1 and int index 5 value 1.
1248        let mut cache_key = name.to_string();
1249        for &(index, value) in bool_constants {
1250            cache_key.push('|');
1251            cache_key.push_str(&index.to_string());
1252            cache_key.push_str(if value { ":b1" } else { ":b0" });
1253        }
1254        for &(index, value) in int_constants {
1255            cache_key.push('|');
1256            cache_key.push_str(&index.to_string());
1257            cache_key.push(':');
1258            cache_key.push('i');
1259            cache_key.push_str(&value.to_string());
1260        }
1261
1262        if !self.cache.contains_key(&cache_key) {
1263            // Build the FunctionConstantValues object with all bool and i32
1264            // constants.  Metal's set_constant_value_at_index reads the value
1265            // through a raw pointer; the pointed-to bytes must match the size
1266            // declared in the MSL shader (1 byte for bool, 4 bytes for int).
1267            let fcv = FunctionConstantValues::new();
1268
1269            for &(index, value) in bool_constants {
1270                // MTLDataType::Bool = 53 (metal-rs argument.rs).
1271                // The Metal runtime reads it as an Objective-C BOOL (uint8_t).
1272                let v: u8 = if value { 1 } else { 0 };
1273                fcv.set_constant_value_at_index(
1274                    (&v as *const u8).cast::<std::ffi::c_void>(),
1275                    MTLDataType::Bool,
1276                    index as u64,
1277                );
1278            }
1279
1280            for &(index, value) in int_constants {
1281                // MTLDataType::Int = 29 (metal-rs argument.rs).
1282                // The Metal runtime reads 4 bytes as a signed 32-bit integer,
1283                // matching the Metal shader type `constant int`.
1284                fcv.set_constant_value_at_index(
1285                    (&value as *const i32).cast::<std::ffi::c_void>(),
1286                    MTLDataType::Int,
1287                    index as u64,
1288                );
1289            }
1290
1291            // ADR-029 iter-175 Step 1l: try precompiled .metallib first.
1292            // Step 1m: gated separately on MLX_PRECOMPILED_METALLIB_FCV=1
1293            // so we can isolate whether the integration regression at
1294            // Step 1l (tg50 95.4 → 62.1) lives in this FCV path vs the
1295            // no-FCV path in `get_pipeline`.
1296            //
1297            // get_function with FCV takes ownership of the FCV, so we
1298            // build a separate one for the precompiled probe.
1299            let precompiled_function = if precompiled_fcv_enabled() {
1300                let probe_fcv = FunctionConstantValues::new();
1301                for &(index, value) in bool_constants {
1302                    let v: u8 = if value { 1 } else { 0 };
1303                    probe_fcv.set_constant_value_at_index(
1304                        (&v as *const u8).cast::<std::ffi::c_void>(),
1305                        MTLDataType::Bool,
1306                        index as u64,
1307                    );
1308                }
1309                for &(index, value) in int_constants {
1310                    probe_fcv.set_constant_value_at_index(
1311                        (&value as *const i32).cast::<std::ffi::c_void>(),
1312                        MTLDataType::Int,
1313                        index as u64,
1314                    );
1315                }
1316                self.try_precompiled_lib(device)
1317                    .and_then(|lib| lib.get_function(name, Some(probe_fcv)).ok())
1318            } else {
1319                None
1320            };
1321
1322            let function = match precompiled_function {
1323                Some(f) => f,
1324                None => {
1325                    // Slow path: compile the shader with function constant specialisation.
1326                    let source = self.sources.get(name).ok_or_else(|| {
1327                        MlxError::KernelNotFound(name.to_string())
1328                    })?;
1329
1330                    let compile_opts = metal::CompileOptions::new();
1331                    let library = device
1332                        .new_library_with_source(source, &compile_opts)
1333                        .map_err(|msg| MlxError::ShaderCompilationError {
1334                            name: name.to_string(),
1335                            message: msg,
1336                        })?;
1337
1338                    library
1339                        .get_function(name, Some(fcv))
1340                        .map_err(|msg| MlxError::ShaderCompilationError {
1341                            name: name.to_string(),
1342                            message: msg,
1343                        })?
1344                }
1345            };
1346
1347            // Label this specialisation with the full composite cache key
1348            // (e.g. `kernel_mul_mv_q4_0_f32|0:b1|3:i32`) so xctrace Metal
1349            // System Trace shows each function-constant variant as a distinct
1350            // pipeline.  Without this, all specialisations share a generic
1351            // "Compute Command 0" identifier and we cannot attribute µs/token
1352            // to a specific (kernel, constants) combination.  ADR-015 iter9b.
1353            let descriptor = ComputePipelineDescriptor::new();
1354            descriptor.set_compute_function(Some(&function));
1355            descriptor.set_label(&cache_key);
1356            // ADR-028 iter-376: same hint as primary pipeline path.
1357            if std::env::var("HF2Q_PIPELINE_TG_MULT_HINT").ok().as_deref() == Some("1") {
1358                descriptor.set_thread_group_size_is_multiple_of_thread_execution_width(true);
1359            }
1360
1361            let pipeline = device
1362                .new_compute_pipeline_state(&descriptor)
1363                .map_err(|msg| MlxError::ShaderCompilationError {
1364                    name: name.to_string(),
1365                    message: msg,
1366                })?;
1367
1368            self.cache.insert(cache_key.clone(), pipeline);
1369        }
1370
1371        self.cache.get(&cache_key).ok_or_else(|| {
1372            MlxError::KernelNotFound(name.to_string())
1373        })
1374    }
1375
1376    /// Get a compiled compute pipeline for the named kernel, specialized with
1377    /// Metal bool function constants.
1378    ///
1379    /// The `bool_constants` slice contains `(index, value)` pairs.  Each pair
1380    /// maps to a `[[function_constant(index)]]` declaration in the MSL shader.
1381    ///
1382    /// This is a thin wrapper around [`get_pipeline_with_constants`] that
1383    /// passes an empty `int_constants` slice.  Existing callers continue to
1384    /// work without modification; the cache-key format for pure-bool pipelines
1385    /// is compatible (bool entries carry the 'b' type marker, which is the
1386    /// only format ever written by this wrapper).
1387    ///
1388    /// # Errors
1389    ///
1390    /// * `MlxError::KernelNotFound` — no source registered for this name.
1391    /// * `MlxError::ShaderCompilationError` — MSL compilation, function
1392    ///   specialisation, or pipeline creation failed.
1393    pub fn get_pipeline_with_bool_constants(
1394        &mut self,
1395        name: &str,
1396        device: &metal::DeviceRef,
1397        bool_constants: &[(usize, bool)],
1398    ) -> Result<&ComputePipelineState> {
1399        self.get_pipeline_with_constants(name, device, bool_constants, &[])
1400    }
1401
1402    /// Check if a pipeline for the given name is already compiled and cached.
1403    pub fn is_cached(&self, name: &str) -> bool {
1404        self.cache.contains_key(name)
1405    }
1406
1407    /// Number of compiled pipelines currently in the cache.
1408    pub fn cached_count(&self) -> usize {
1409        self.cache.len()
1410    }
1411
1412    /// Number of registered shader sources.
1413    pub fn source_count(&self) -> usize {
1414        self.sources.len()
1415    }
1416}
1417
1418impl Default for KernelRegistry {
1419    fn default() -> Self {
1420        Self::new()
1421    }
1422}
1423
1424#[cfg(test)]
1425mod tests {
1426    use super::*;
1427
1428    /// Minimal Metal shader that uses a single int function constant.
1429    ///
1430    /// The kernel writes the constant value N into the first element of the
1431    /// output buffer, allowing the test to verify that the Metal compiler
1432    /// actually sees distinct specialisations for N=4 and N=8.
1433    ///
1434    /// The shader is intentionally trivial — we only need it to *compile* with
1435    /// an int function constant; correctness of the kernel logic is not under
1436    /// test here.
1437    const INT_FC_TEST_SHADER: &str = r#"
1438#include <metal_stdlib>
1439using namespace metal;
1440
1441constant int test_N [[function_constant(100)]];
1442
1443kernel void int_fc_test_kernel(
1444    device int* out [[buffer(0)]],
1445    uint tid [[thread_position_in_grid]])
1446{
1447    if (tid == 0) {
1448        out[0] = test_N;
1449    }
1450}
1451"#;
1452
1453    /// Verify that `get_pipeline_with_constants` produces distinct cached
1454    /// pipelines for different i32 function-constant values, and that
1455    /// `get_pipeline_with_bool_constants` (the backward-compat wrapper) still
1456    /// works correctly with the new 'b'-prefixed cache-key format.
1457    ///
1458    /// This test requires a real Metal device and is therefore marked
1459    /// `#[ignore]` on non-Apple platforms, but runs unconditionally on macOS.
1460    #[test]
1461    fn test_int_fc_distinct_pipelines_and_bool_compat() {
1462        let device = metal::Device::system_default()
1463            .expect("no Metal device — run on Apple Silicon or x86 Mac with Metal support");
1464
1465        let mut registry = KernelRegistry::new();
1466
1467        // Register the inline test shader under a name that cannot collide with
1468        // any production kernel.
1469        registry.register_source("int_fc_test_kernel", INT_FC_TEST_SHADER);
1470
1471        // Compile with N=4.
1472        let p4_ptr = registry
1473            .get_pipeline_with_constants(
1474                "int_fc_test_kernel",
1475                &device,
1476                &[],                  // no bool constants
1477                &[(100, 4_i32)],      // int constant index 100 = 4
1478            )
1479            .expect("pipeline N=4 should compile") as *const _;
1480
1481        // Cache must now have exactly 1 entry for this kernel.
1482        // (Other production kernels may already be in cache from new(); here
1483        // we check that the N=4 key was inserted.)
1484        let count_after_n4 = registry.cached_count();
1485
1486        // Compile with N=8 — must produce a SEPARATE pipeline.
1487        let p8_ptr = registry
1488            .get_pipeline_with_constants(
1489                "int_fc_test_kernel",
1490                &device,
1491                &[],
1492                &[(100, 8_i32)],
1493            )
1494            .expect("pipeline N=8 should compile") as *const _;
1495
1496        // Cache must have grown by exactly 1.
1497        assert_eq!(
1498            registry.cached_count(),
1499            count_after_n4 + 1,
1500            "N=8 must produce a new cache entry"
1501        );
1502
1503        // The two pipelines must be distinct objects in the cache.
1504        assert_ne!(
1505            p4_ptr, p8_ptr,
1506            "N=4 and N=8 specialisations must be separate ComputePipelineState objects"
1507        );
1508
1509        // A second call with N=4 must return the SAME pipeline (cache hit, no
1510        // new compilation).
1511        let p4_again_ptr = registry
1512            .get_pipeline_with_constants(
1513                "int_fc_test_kernel",
1514                &device,
1515                &[],
1516                &[(100, 4_i32)],
1517            )
1518            .expect("pipeline N=4 cache hit should succeed") as *const _;
1519
1520        assert_eq!(
1521            registry.cached_count(),
1522            count_after_n4 + 1,
1523            "repeated N=4 call must be a cache hit, not a new entry"
1524        );
1525        assert_eq!(
1526            p4_ptr, p4_again_ptr,
1527            "repeated N=4 call must return the same pipeline pointer"
1528        );
1529
1530        // Verify backward compatibility: get_pipeline_with_bool_constants must
1531        // still route through get_pipeline_with_constants and produce a cached
1532        // pipeline without panicking.
1533        //
1534        // We register a separate bool-constant shader that does NOT use a bool
1535        // function constant (so the Metal compiler ignores missing FCs for
1536        // this trivial case) — but the call path and cache-key format are what
1537        // matter here.  We reuse the int_fc_test_kernel source; the bool FC is
1538        // simply unused by the shader (Metal allows unused FCs when the shader
1539        // declares them with `function_constant` but the value is never read).
1540        //
1541        // To avoid a Metal compiler error for an undeclared function constant,
1542        // we register a separate bare-kernel shader for the bool wrapper test.
1543        const BARE_SHADER: &str = r#"
1544#include <metal_stdlib>
1545using namespace metal;
1546kernel void bare_kernel(device int* out [[buffer(0)]], uint tid [[thread_position_in_grid]]) {
1547    if (tid == 0) { out[0] = 42; }
1548}
1549"#;
1550        registry.register_source("bare_kernel", BARE_SHADER);
1551
1552        let count_before_bool = registry.cached_count();
1553        let _bool_pipeline = registry
1554            .get_pipeline_with_bool_constants("bare_kernel", &device, &[])
1555            .expect("bool-constants wrapper with empty slice must succeed");
1556
1557        assert_eq!(
1558            registry.cached_count(),
1559            count_before_bool + 1,
1560            "bool-constants wrapper must insert one new cache entry"
1561        );
1562    }
1563
1564    /// Verify that the `MTLComputePipelineState.label` produced by
1565    /// `get_pipeline` and `get_pipeline_with_constants` actually propagates
1566    /// from the descriptor to the resulting pipeline state.
1567    ///
1568    /// This is the in-process smoke check for ADR-015 iter9b: we cannot
1569    /// reach into xctrace from Rust, but we can read back the same `label`
1570    /// property xctrace consumes via `ComputePipelineStateRef::label()`.
1571    /// If labels are missing or wrong here, the MST trace will also show
1572    /// generic identifiers — so this test gates the iter9 retry's
1573    /// per-Q4_0-kernel attribution.
1574    #[test]
1575    fn test_pipeline_labels_propagate_for_mst() {
1576        let device = metal::Device::system_default()
1577            .expect("no Metal device — run on Apple Silicon or x86 Mac with Metal support");
1578
1579        let mut registry = KernelRegistry::new();
1580
1581        // Reuse the same trivial shaders as the int-FC test.
1582        registry.register_source("int_fc_test_kernel", INT_FC_TEST_SHADER);
1583
1584        const BARE_SHADER_LABEL_TEST: &str = r#"
1585#include <metal_stdlib>
1586using namespace metal;
1587kernel void label_smoke_kernel(device int* out [[buffer(0)]], uint tid [[thread_position_in_grid]]) {
1588    if (tid == 0) { out[0] = 7; }
1589}
1590"#;
1591        registry.register_source("label_smoke_kernel", BARE_SHADER_LABEL_TEST);
1592
1593        // Plain get_pipeline path — label must equal the kernel name.
1594        // Capture as owned String so the cache borrow is released before
1595        // the next get_pipeline_with_constants call below.
1596        let plain_label = registry
1597            .get_pipeline("label_smoke_kernel", &device)
1598            .expect("plain pipeline must compile")
1599            .label()
1600            .to_string();
1601        assert_eq!(
1602            plain_label, "label_smoke_kernel",
1603            "get_pipeline must label the pipeline with the kernel name (xctrace MST attribution)"
1604        );
1605
1606        // Constants path — label must equal the composite cache key so each
1607        // function-constant variant is individually attributable in MST.
1608        // We capture the label as an owned String to release the borrow on
1609        // the cache before fetching the next specialisation.
1610        let label_v7 = registry
1611            .get_pipeline_with_constants(
1612                "int_fc_test_kernel",
1613                &device,
1614                &[],
1615                &[(100, 7_i32)],
1616            )
1617            .expect("specialised pipeline must compile")
1618            .label()
1619            .to_string();
1620        assert_eq!(
1621            label_v7, "int_fc_test_kernel|100:i7",
1622            "get_pipeline_with_constants must label with the cache_key so each \
1623             specialisation is distinct in xctrace MST"
1624        );
1625
1626        // A second specialisation must produce a different label.
1627        let label_v13 = registry
1628            .get_pipeline_with_constants(
1629                "int_fc_test_kernel",
1630                &device,
1631                &[],
1632                &[(100, 13_i32)],
1633            )
1634            .expect("second specialised pipeline must compile")
1635            .label()
1636            .to_string();
1637        assert_eq!(label_v13, "int_fc_test_kernel|100:i13");
1638        assert_ne!(
1639            label_v7, label_v13,
1640            "distinct constant values must yield distinct pipeline labels"
1641        );
1642    }
1643}