mlx_native/ops/rope_multi.rs
1//! Multi-section Rotary Position Embedding with optional interleaved mode.
2//!
3//! Used by Qwen3.5 / Qwen3.6 full-attention layers (ADR-013 Decision 10).
4//! Both MROPE (`mode = 8`) and IMROPE (`mode = 40`) share a kernel; only the
5//! sector-to-axis mapping differs.
6//!
7//! # Spec (summary)
8//!
9//! For every pair `p ∈ [0, rope_dim/2)`:
10//! 1. `sector = p mod (s0 + s1 + s2 + s3)`
11//! 2. Pick axis based on `mode`:
12//! * `Mrope`: contiguous sections — `sector ∈ [0, s0)` → axis 0, etc.
13//! * `Imrope`: `sector % 3` cycling — `sector % 3 == 0 && sector < 3*s0`
14//! → axis 0; `== 1 && sector < 3*s1` → axis 1; `== 2 && sector < 3*s2`
15//! → axis 2; else axis 3.
16//! 3. `theta = position[axis] * freq_base^(-2p/rope_dim)`
17//! 4. Rotate pair `(x[p], x[p + head_dim/2])` by that angle (NeoX indexing).
18//!
19//! Pairs `p ≥ rope_dim/2` pass through unchanged (partial-rotary-factor).
20//!
21//! # Positions layout
22//!
23//! The `positions` buffer is an `int32` array of length `4 * seq_len`:
24//! first `seq_len` entries are the time-axis positions, next `seq_len` are
25//! the height-axis, then width, then extra. For Qwen3.5 text, all four
26//! axes are set to the token's 1D position.
27
28use std::cell::RefCell;
29use std::collections::HashMap;
30
31use metal::MTLSize;
32
33use crate::buffer::MlxBuffer;
34use crate::dtypes::DType;
35use crate::encoder::CommandEncoder;
36use crate::error::{MlxError, Result};
37use crate::kernel_registry::KernelRegistry;
38
39pub static ROPE_MULTI_SHADER_SOURCE: &str = include_str!("../shaders/rope_multi.metal");
40
41pub fn register(registry: &mut KernelRegistry) {
42 registry.register_source("rope_multi_f32", ROPE_MULTI_SHADER_SOURCE);
43 registry.register_source("rope_multi_bf16", ROPE_MULTI_SHADER_SOURCE);
44}
45
46/// MROPE variant. Wire-level values match the ggml `GGML_ROPE_TYPE_*` enum.
47#[derive(Debug, Clone, Copy, PartialEq, Eq)]
48#[repr(u32)]
49pub enum RopeMultiMode {
50 /// Standard multi-section RoPE; contiguous sections.
51 Mrope = 8,
52 /// Interleaved multi-section RoPE; `sector % 3` cycles through 3 axes.
53 /// Used by Qwen3.5 / Qwen3.6.
54 Imrope = 40,
55 /// Vision multi-section RoPE for ViT 2-D positions (Qwen3-VL ViT block).
56 ///
57 /// Mode value `24` matches `GGML_ROPE_TYPE_VISION` in
58 /// `/opt/llama.cpp/ggml/include/ggml.h:253` and the per-section
59 /// `[yyyyxxxx]` layout described at `ggml.h:1840-1846`.
60 ///
61 /// # Layout
62 ///
63 /// Only the first two section counts (`s0 = y`, `s1 = x`) are used; the
64 /// last two are ignored. With `n_dims = head_dim / 2` and `sect_dims =
65 /// s0 + s1 = n_dims`, the rotated pairs partition as:
66 ///
67 /// ```text
68 /// pair_idx in [0, s0) -> axis 0 (y), local_p = pair_idx
69 /// pair_idx in [s0, s0 + s1) -> axis 1 (x), local_p = pair_idx - s0
70 /// ```
71 ///
72 /// # Per-section theta
73 ///
74 /// Unlike `Mrope` / `Imrope` which use a unified theta sequence across
75 /// all sections (`theta = pos[axis] * freq_base^(-2*pair_idx/rope_dim)`),
76 /// vision rope **restarts the theta exponent at every section boundary**:
77 ///
78 /// ```text
79 /// theta_scale = freq_base^(-2 / n_dims) where n_dims = head_dim/2
80 /// theta = pos[axis] * theta_scale^local_p
81 /// ```
82 ///
83 /// `local_p` is the index of the pair *within its section*, not the
84 /// global `pair_idx`. This per-section restart is what produces the
85 /// `[0123][0123]` exponent pattern documented at `ggml.h:1845-1846`.
86 ///
87 /// # No partial-rotary tail
88 ///
89 /// The CPU reference at `ggml-cpu/ops.cpp:5860` calls
90 /// `rotate_pairs(ne0, n_dims, ...)` (rotating *all* `head_dim/2` pairs)
91 /// and at `:5866` skips the partial-rotary fill loop when `is_vision`,
92 /// so the caller MUST supply `rope_dim == head_dim`.
93 ///
94 /// # Caller-side requirements
95 ///
96 /// - `rope_dim == head_dim` (validated; no partial-rotary support).
97 /// - `sections[0] + sections[1] == head_dim / 2` (validated; the last
98 /// two sections are ignored but must be present in the buffer for
99 /// binary-layout uniformity with `Mrope` / `Imrope`).
100 /// - `positions` buffer length still `4 * seq_len`; only axes 0 and 1
101 /// are read.
102 Vision = 24,
103}
104
105/// Shape + config for a rope_multi dispatch.
106#[derive(Debug, Clone, Copy)]
107pub struct RopeMultiParams {
108 pub head_dim: u32,
109 pub rope_dim: u32, // must be <= head_dim; must be even
110 pub n_heads: u32,
111 pub seq_len: u32,
112 pub freq_base: f32,
113 pub mode: RopeMultiMode,
114 /// Section counts `[s0, s1, s2, s3]`. Sum should be `rope_dim / 2` for
115 /// full coverage; the kernel tolerates smaller sums (sector wraps).
116 pub sections: [u32; 4],
117}
118
119fn validate(
120 p: &RopeMultiParams,
121 input: &MlxBuffer,
122 output: &MlxBuffer,
123 positions: &MlxBuffer,
124) -> Result<()> {
125 if p.head_dim == 0 || p.rope_dim == 0 || p.n_heads == 0 || p.seq_len == 0 {
126 return Err(MlxError::InvalidArgument(
127 "rope_multi: head_dim, rope_dim, n_heads, seq_len must all be > 0".into(),
128 ));
129 }
130 if p.head_dim % 2 != 0 || p.rope_dim % 2 != 0 {
131 return Err(MlxError::InvalidArgument(
132 "rope_multi: head_dim and rope_dim must be even".into(),
133 ));
134 }
135 if p.rope_dim > p.head_dim {
136 return Err(MlxError::InvalidArgument(
137 "rope_multi: rope_dim must be <= head_dim".into(),
138 ));
139 }
140 if !p.freq_base.is_finite() || p.freq_base <= 0.0 {
141 return Err(MlxError::InvalidArgument(format!(
142 "rope_multi: freq_base must be finite and positive, got {}",
143 p.freq_base
144 )));
145 }
146 // Vision-mode requires every pair to rotate (no partial-rotary tail —
147 // see /opt/llama.cpp/ggml/src/ggml-cpu/ops.cpp:5803,5866) and the first
148 // two section counts must sum to n_dims = head_dim / 2 (last 2 ignored
149 // per ggml.h:1843-1846).
150 if p.mode == RopeMultiMode::Vision {
151 if p.rope_dim != p.head_dim {
152 return Err(MlxError::InvalidArgument(format!(
153 "rope_multi(Vision): rope_dim must equal head_dim (no partial-rotary tail in vision mode), got rope_dim={}, head_dim={}",
154 p.rope_dim, p.head_dim
155 )));
156 }
157 let n_dims = p.head_dim / 2;
158 let sect_sum = p.sections[0] + p.sections[1];
159 if sect_sum != n_dims {
160 return Err(MlxError::InvalidArgument(format!(
161 "rope_multi(Vision): sections[0] + sections[1] must equal head_dim/2 ({}), got {} + {} = {}",
162 n_dims, p.sections[0], p.sections[1], sect_sum
163 )));
164 }
165 }
166
167 let n_rows = (p.seq_len as usize) * (p.n_heads as usize);
168 let elements = n_rows * (p.head_dim as usize);
169 if input.element_count() != elements {
170 return Err(MlxError::InvalidArgument(format!(
171 "rope_multi: input element count {} != seq_len({}) * n_heads({}) * head_dim({}) = {}",
172 input.element_count(),
173 p.seq_len,
174 p.n_heads,
175 p.head_dim,
176 elements
177 )));
178 }
179 if output.element_count() != elements {
180 return Err(MlxError::InvalidArgument(format!(
181 "rope_multi: output element count {} != {}",
182 output.element_count(),
183 elements
184 )));
185 }
186 if input.dtype() != output.dtype() {
187 return Err(MlxError::InvalidArgument(format!(
188 "rope_multi: input/output dtype mismatch {} vs {}",
189 input.dtype(),
190 output.dtype()
191 )));
192 }
193
194 let expected_positions = 4 * (p.seq_len as usize);
195 if positions.element_count() != expected_positions {
196 return Err(MlxError::InvalidArgument(format!(
197 "rope_multi: positions length {} != 4 * seq_len({}) = {}",
198 positions.element_count(),
199 p.seq_len,
200 expected_positions
201 )));
202 }
203 match positions.dtype() {
204 DType::I32 | DType::U32 => {}
205 other => {
206 return Err(MlxError::InvalidArgument(format!(
207 "rope_multi: positions must be i32 or u32 (got {})",
208 other
209 )));
210 }
211 }
212
213 Ok(())
214}
215
216/// Dispatch a rope_multi operation.
217///
218/// The caller must upload:
219/// - `params_buf`: float4 `[freq_base, head_dim, rope_dim, 0]`.
220/// - `rope_params_buf`: uint4 `[n_heads, mode_code, seq_len, 0]`. The
221/// `mode_code` is the `u32` underlying [`RopeMultiMode`].
222/// - `sections_buf`: uint4 `[s0, s1, s2, s3]`.
223/// - `positions`: int32 array of length `4 * seq_len`.
224///
225/// The helper [`build_rope_multi_buffers`] constructs all three small buffers
226/// in one call for callers that do not already keep them pooled.
227#[allow(clippy::too_many_arguments)]
228pub fn dispatch_rope_multi(
229 encoder: &mut CommandEncoder,
230 registry: &mut KernelRegistry,
231 device: &metal::DeviceRef,
232 input: &MlxBuffer,
233 output: &MlxBuffer,
234 positions: &MlxBuffer,
235 params_buf: &MlxBuffer,
236 rope_params_buf: &MlxBuffer,
237 sections_buf: &MlxBuffer,
238 p: RopeMultiParams,
239) -> Result<()> {
240 validate(&p, input, output, positions)?;
241
242 let kernel_name = match input.dtype() {
243 DType::F32 => "rope_multi_f32",
244 DType::BF16 => "rope_multi_bf16",
245 other => {
246 return Err(MlxError::InvalidArgument(format!(
247 "rope_multi: unsupported dtype {}",
248 other
249 )));
250 }
251 };
252
253 let pipeline = registry.get_pipeline(kernel_name, device)?;
254
255 let half_dim = p.head_dim / 2;
256 let n_rows = p.seq_len * p.n_heads;
257
258 // Grid: (half_dim, n_rows). Every thread writes a NeoX pair.
259 let grid = MTLSize::new(half_dim as u64, n_rows as u64, 1);
260
261 let tg_x = std::cmp::min(half_dim, 256).max(1);
262 let remain = (256u32 / tg_x).max(1);
263 let tg_y = std::cmp::min(n_rows, remain).max(1);
264 let tg = MTLSize::new(tg_x as u64, tg_y as u64, 1);
265
266 encoder.encode(
267 pipeline,
268 &[
269 (0, input),
270 (1, output),
271 (2, params_buf),
272 (3, positions),
273 (4, rope_params_buf),
274 (5, sections_buf),
275 ],
276 grid,
277 tg,
278 );
279
280 Ok(())
281}
282
283/// Pre-built triple of small parameter buffers for a `rope_multi` dispatch.
284///
285/// Held in the per-thread [`ROPE_PACK_CACHE`] so callers that issue
286/// repeated dispatches with stable shape (the qwen35 / qwen36 decode hot
287/// path: identical `head_dim`, `rope_dim`, `n_heads`, `seq_len=1`,
288/// `freq_base`, `mode`, `sections` every step) skip the per-call
289/// allocation triplet (~208 µs/token measured on M5 Max in
290/// `cfa-20260426-adr015-wave2a-p3aprime`). Decode-out-of-scope cases
291/// (variable `seq_len`) populate one entry per `seq_len` value seen,
292/// then reuse on re-encounter.
293pub struct RopeMultiBufferPack {
294 pub params_buf: MlxBuffer,
295 pub rope_params_buf: MlxBuffer,
296 pub sections_buf: MlxBuffer,
297}
298
299/// Cache key for [`ROPE_PACK_CACHE`]. Includes the [`MlxDevice`] pointer
300/// so two consecutive sessions with different devices (e.g. model swap)
301/// never share entries — required for correctness, not just isolation.
302#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
303struct RopeMultiCacheKey {
304 device_ptr: usize,
305 head_dim: u32,
306 rope_dim: u32,
307 n_heads: u32,
308 seq_len: u32,
309 freq_base_bits: u32,
310 mode: u32,
311 sections: [u32; 4],
312}
313
314impl RopeMultiCacheKey {
315 fn from_params(device: &crate::MlxDevice, p: &RopeMultiParams) -> Self {
316 Self {
317 device_ptr: device as *const _ as usize,
318 head_dim: p.head_dim,
319 rope_dim: p.rope_dim,
320 n_heads: p.n_heads,
321 seq_len: p.seq_len,
322 freq_base_bits: p.freq_base.to_bits(),
323 mode: p.mode as u32,
324 sections: p.sections,
325 }
326 }
327}
328
329thread_local! {
330 /// Per-thread cache of pre-built `rope_multi` parameter buffers,
331 /// keyed by [`RopeMultiCacheKey`]. Built lazily on first
332 /// [`dispatch_rope_multi_cached`] call for a given key, retained for
333 /// the thread's lifetime (cleared by [`clear_rope_pack_cache`] in
334 /// tests / on explicit model unload).
335 static ROPE_PACK_CACHE: RefCell<HashMap<RopeMultiCacheKey, RopeMultiBufferPack>> =
336 RefCell::new(HashMap::new());
337}
338
339/// Clear the thread-local rope-multi pack cache.
340///
341/// Useful between model loads (the cache key includes the device pointer
342/// so old entries can never be returned for a new device, but they leak
343/// memory until cleared) and in test suites that swap mocked devices.
344pub fn clear_rope_pack_cache() {
345 ROPE_PACK_CACHE.with(|cell| cell.borrow_mut().clear());
346}
347
348/// Inspect the current pack-cache size — diagnostic only.
349pub fn rope_pack_cache_len() -> usize {
350 ROPE_PACK_CACHE.with(|cell| cell.borrow().len())
351}
352
353/// Dispatch a `rope_multi` operation, reusing pre-built parameter
354/// buffers from the per-thread cache.
355///
356/// Functionally equivalent to [`dispatch_rope_multi`] preceded by
357/// [`build_rope_multi_buffers`], but the small param/rope_params/sections
358/// buffers (3 × 16 bytes each) are built once per
359/// `(device, head_dim, rope_dim, n_heads, seq_len, freq_base, mode,
360/// sections)` tuple and reused on every subsequent call. See
361/// `docs/ADR-015-mlx-native-single-cb-decode.md` §"P3a' live profile pass"
362/// rank-4 finding — `build_rope_multi_buffers` per-call alloc was
363/// measured at 208 µs/token on the qwen3.6-27b-dwq46 dense-FFN-Q hot
364/// path (16 FullAttn layers × 2 calls/layer = 32 calls/token, each
365/// allocating 3 fresh `MlxBuffer`s via Mach IPC). This helper closes
366/// that residual.
367///
368/// Bit-exact to the per-call form: identical kernel, identical inputs,
369/// only the small parameter triplet is sourced from the cache.
370#[allow(clippy::too_many_arguments)]
371pub fn dispatch_rope_multi_cached(
372 encoder: &mut CommandEncoder,
373 registry: &mut KernelRegistry,
374 device: &crate::MlxDevice,
375 input: &MlxBuffer,
376 output: &MlxBuffer,
377 positions: &MlxBuffer,
378 p: RopeMultiParams,
379) -> Result<()> {
380 let key = RopeMultiCacheKey::from_params(device, &p);
381 ROPE_PACK_CACHE.with(|cell| {
382 let mut map = cell.borrow_mut();
383 if !map.contains_key(&key) {
384 let (params_buf, rope_params_buf, sections_buf) =
385 build_rope_multi_buffers(device, p)?;
386 map.insert(
387 key,
388 RopeMultiBufferPack {
389 params_buf,
390 rope_params_buf,
391 sections_buf,
392 },
393 );
394 }
395 let pack = map
396 .get(&key)
397 .expect("inserted above if missing; cache is single-threaded");
398 dispatch_rope_multi(
399 encoder,
400 registry,
401 device.metal_device(),
402 input,
403 output,
404 positions,
405 &pack.params_buf,
406 &pack.rope_params_buf,
407 &pack.sections_buf,
408 p,
409 )
410 })
411}
412
413/// Convenience: build all three small parameter buffers given a [`RopeMultiParams`].
414///
415/// Returns `(params_buf, rope_params_buf, sections_buf)`.
416pub fn build_rope_multi_buffers(
417 device: &crate::MlxDevice,
418 p: RopeMultiParams,
419) -> Result<(MlxBuffer, MlxBuffer, MlxBuffer)> {
420 let mut params = device.alloc_buffer(4 * 4, DType::F32, vec![4])?;
421 {
422 let s = params.as_mut_slice::<f32>()?;
423 s[0] = p.freq_base;
424 s[1] = p.head_dim as f32;
425 s[2] = p.rope_dim as f32;
426 s[3] = 0.0;
427 }
428 let mut rope_params = device.alloc_buffer(4 * 4, DType::U32, vec![4])?;
429 {
430 let s = rope_params.as_mut_slice::<u32>()?;
431 s[0] = p.n_heads;
432 s[1] = p.mode as u32;
433 s[2] = p.seq_len;
434 s[3] = 0;
435 }
436 let mut sections = device.alloc_buffer(4 * 4, DType::U32, vec![4])?;
437 {
438 let s = sections.as_mut_slice::<u32>()?;
439 s[0] = p.sections[0];
440 s[1] = p.sections[1];
441 s[2] = p.sections[2];
442 s[3] = p.sections[3];
443 }
444 Ok((params, rope_params, sections))
445}