File size: 24,406 Bytes
6ed7e38
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
// CUDA-accelerated causal TTT (LaCT fast-weight SwiGLU) forward for VLANeXt.
//
// Strategy: the GEMMs (k@w0, k@w2, v@w1^T, q@w0, q@w2, o@w1, and the three
// grad outer-products k^T@..) are batched [B, *, *] matmuls -- we leave those
// to ATen bmm (cuBLAS). The custom CUDA kernels fuse the launch-storm of tiny
// elementwise / reduction ops:
//   - silu_glu:     hidden = silu(gate) * up
//   - silu_bwd_glu: the dgate/dhidden chain (silu_backprop) for the grad path
//   - frob_normalize_add: w <- w + grad/||grad||_F ; then weight-norm rescale
//
// Parity contract (muon_update_steps == 0): zeropower_via_newtonschulz5 with
// steps=0 reduces to grad / (||grad||_F + 1e-7) per [d,d] head matrix. The
// weight-norm step then rescales each column back to the detached init norm.

#include <torch/extension.h>
#include <vector>

// ---- declarations of kernels defined in ttt_fused.cu ----
namespace ttt_cuda {

// hidden = silu(gate) * up        (all [N, D] contiguous, bf16/fp16/fp32)
torch::Tensor silu_glu(const torch::Tensor& gate, const torch::Tensor& up);

// Given dhidden, gate, up: returns (dgate_before_act, dhidden_before_mul)
// dhidden_before_mul = dhidden * silu(gate)
// dgate_before_act   = silu_backprop(dhidden * up, gate)
std::vector<torch::Tensor> silu_bwd_glu(
    const torch::Tensor& dhidden,
    const torch::Tensor& gate,
    const torch::Tensor& up);

// w_new = (w + grad/(||grad||_F + 1e-7)); then column-normalize to w_init_norm.
// grad is normalized per batch-matrix over dims (1,2). norm_dim selects the
// weight-norm reduction dim (1 for these [B,d,dh] layouts, matching torch).
torch::Tensor frob_norm_update(
    const torch::Tensor& w,
    const torch::Tensor& grad,
    const torch::Tensor& w_init_norm,
    int64_t norm_dim);

// backward primitives
std::vector<torch::Tensor> silu_derivs(const torch::Tensor& x);   // returns {silu', silu''}
torch::Tensor frobnorm_bwd(const torch::Tensor& gy, const torch::Tensor& x, double eps);
torch::Tensor weightnorm_bwd(const torch::Tensor& gy, const torch::Tensor& w_pre,
                             const torch::Tensor& wn_target, double eps);
torch::Tensor infer_step(const torch::Tensor& q, const torch::Tensor& w0,
                         const torch::Tensor& w2, const torch::Tensor& w1,
                         const torch::Tensor& o_norm_weight, double eps);
torch::Tensor infer_step_mid(const torch::Tensor& q, const torch::Tensor& w0,
                             const torch::Tensor& w2, const torch::Tensor& w1,
                             const torch::Tensor& o_norm_weight, double eps, double qeps);

}  // namespace ttt_cuda

// ---- helpers (host-side, ATen) ----
static inline torch::Tensor frob_normalize(const torch::Tensor& g) {
    // grad / (||grad||_F + 1e-7), Frobenius over last two dims, per batch.
    auto nrm = g.flatten(1).norm(2, /*dim=*/1, /*keepdim=*/true).unsqueeze(-1);
    return g / (nrm + 1e-7);
}

// One fast-weight update step (apply-then-update uses this for the update half).
// Mutates w0,w1,w2 in place (returns new tensors). steps==0 parity path.
static void fw_update(
    torch::Tensor& w0, torch::Tensor& w1, torch::Tensor& w2,
    const torch::Tensor& ki, const torch::Tensor& vi,
    const torch::Tensor& lr0i, const torch::Tensor& lr1i, const torch::Tensor& lr2i,
    const torch::Tensor& w0_norm, const torch::Tensor& w1_norm, const torch::Tensor& w2_norm) {

    auto gate = ki.bmm(w0);              // [B, l, dh]
    auto up   = ki.bmm(w2);              // [B, l, dh]
    auto hidden = ttt_cuda::silu_glu(gate, up);

    auto dhidden = vi.bmm(w1.transpose(-1, -2));   // [B, l, dh]
    auto chain = ttt_cuda::silu_bwd_glu(dhidden, gate, up);
    auto dgate_before_act   = chain[0];            // [B, l, dh]
    auto dhidden_before_mul = chain[1];            // [B, l, dh]

    // grads (Frobenius-normalized, steps==0)
    auto w1_grad = frob_normalize(
        (hidden * lr1i).to(vi.dtype()).transpose(-1, -2).bmm(vi));      // [B, dh, d]
    auto w0_grad = frob_normalize(
        (ki * lr0i).to(dgate_before_act.dtype()).transpose(-1, -2).bmm(dgate_before_act)); // [B, d, dh]
    auto w2_grad = frob_normalize(
        (ki * lr2i).to(dhidden_before_mul.dtype()).transpose(-1, -2).bmm(dhidden_before_mul)); // [B, d, dh]

    w1 = ttt_cuda::frob_norm_update(w1, w1_grad, w1_norm, /*norm_dim=*/1);
    w0 = ttt_cuda::frob_norm_update(w0, w0_grad, w0_norm, /*norm_dim=*/1);
    w2 = ttt_cuda::frob_norm_update(w2, w2_grad, w2_norm, /*norm_dim=*/1);
}

// fw_update that ALSO returns the intermediates the backward needs, so chunk_vjp
// can skip recomputing them (Phase 2). Returns, in order:
//   {gate, up, dhidden, fn0_in, fn1_in, fn2_in, w0_pre, w1_pre, w2_pre}
// where fn*_in are the raw (pre-Frobenius) grad outer products and w*_pre are the
// post-(W+frobnorm) weights BEFORE the weight-norm rescale. w0/w1/w2 are updated
// in place to the post-weight-norm new weights (same as fw_update).
static std::array<torch::Tensor, 9> fw_update_save(
    torch::Tensor& w0, torch::Tensor& w1, torch::Tensor& w2,
    const torch::Tensor& ki, const torch::Tensor& vi,
    const torch::Tensor& lr0i, const torch::Tensor& lr1i, const torch::Tensor& lr2i,
    const torch::Tensor& w0_norm, const torch::Tensor& w1_norm, const torch::Tensor& w2_norm) {

    auto gate = ki.bmm(w0);
    auto up   = ki.bmm(w2);
    auto hidden = ttt_cuda::silu_glu(gate, up);
    auto dhidden = vi.bmm(w1.transpose(-1, -2));
    auto chain = ttt_cuda::silu_bwd_glu(dhidden, gate, up);
    auto dgate_before_act   = chain[0];
    auto dhidden_before_mul = chain[1];

    auto fn1_in = (hidden * lr1i).to(vi.dtype()).transpose(-1, -2).bmm(vi);
    auto fn0_in = (ki * lr0i).to(dgate_before_act.dtype()).transpose(-1, -2).bmm(dgate_before_act);
    auto fn2_in = (ki * lr2i).to(dhidden_before_mul.dtype()).transpose(-1, -2).bmm(dhidden_before_mul);
    auto w1_grad = frob_normalize(fn1_in);
    auto w0_grad = frob_normalize(fn0_in);
    auto w2_grad = frob_normalize(fn2_in);

    auto w0_pre = w0 + w0_grad;
    auto w1_pre = w1 + w1_grad;
    auto w2_pre = w2 + w2_grad;

    w1 = ttt_cuda::frob_norm_update(w1, w1_grad, w1_norm, 1);
    w0 = ttt_cuda::frob_norm_update(w0, w0_grad, w0_norm, 1);
    w2 = ttt_cuda::frob_norm_update(w2, w2_grad, w2_norm, 1);

    return {gate, up, dhidden, fn0_in, fn1_in, fn2_in, w0_pre, w1_pre, w2_pre};
}

// output_i = (silu(qi@w0) * (qi@w2)) @ w1
static torch::Tensor fw_apply(
    const torch::Tensor& qi, const torch::Tensor& w0,
    const torch::Tensor& w1, const torch::Tensor& w2) {
    auto gate = qi.bmm(w0);
    auto up   = qi.bmm(w2);
    auto h = ttt_cuda::silu_glu(gate, up);
    return h.bmm(w1);
}

std::vector<torch::Tensor> causal_ttt_forward(
    torch::Tensor w0, torch::Tensor w1, torch::Tensor w2,
    torch::Tensor q, torch::Tensor k, torch::Tensor v,
    torch::Tensor lr0, torch::Tensor lr1, torch::Tensor lr2,
    int64_t chunk_size,
    c10::optional<torch::Tensor> vlm_k,
    c10::optional<torch::Tensor> vlm_v,
    c10::optional<torch::Tensor> vlm_lr0,
    c10::optional<torch::Tensor> vlm_lr1,
    c10::optional<torch::Tensor> vlm_lr2) {

    TORCH_CHECK(q.is_cuda(), "causal_ttt_forward: inputs must be CUDA tensors");

    // detached init column norms (weight-norm targets), dim=1 like torch ref
    auto w0_norm = w0.detach().norm(2, /*dim=*/1, /*keepdim=*/true);
    auto w1_norm = w1.detach().norm(2, /*dim=*/1, /*keepdim=*/true);
    auto w2_norm = w2.detach().norm(2, /*dim=*/1, /*keepdim=*/true);

    // ---- global (non-causal) VLM pre-update: makes VLM fully visible ----
    if (vlm_k.has_value()) {
        fw_update(w0, w1, w2,
                  vlm_k.value(), vlm_v.value(),
                  vlm_lr0.value(), vlm_lr1.value(), vlm_lr2.value(),
                  w0_norm, w1_norm, w2_norm);
    }

    const int64_t L = q.size(1);
    std::vector<torch::Tensor> outs;
    for (int64_t s = 0; s < L; s += chunk_size) {
        int64_t e = std::min(s + chunk_size, L);
        using torch::indexing::Slice;

        // apply current fast weights to this chunk's query (apply-then-update)
        auto qi = q.index({Slice(), Slice(s, e), Slice()});
        outs.push_back(fw_apply(qi, w0, w1, w2));

        // then update with this chunk's (k, v)
        auto ki = k.index({Slice(), Slice(s, e), Slice()});
        auto vi = v.index({Slice(), Slice(s, e), Slice()});
        auto l0 = lr0.index({Slice(), Slice(s, e), Slice()});
        auto l1 = lr1.index({Slice(), Slice(s, e), Slice()});
        auto l2 = lr2.index({Slice(), Slice(s, e), Slice()});
        fw_update(w0, w1, w2, ki, vi, l0, l1, l2, w0_norm, w1_norm, w2_norm);
    }

    auto output = torch::cat(outs, /*dim=*/1);
    return {output, w0, w1, w2};
}

// Forward that ALSO saves per-chunk entry weights (for the no-recompute
// backward, Phase 1). Returns:
//   [output, w0, w1, w2,           (final weights)
//    entry_w0, entry_w1, entry_w2, (stacked [n_chunk, B, *, *] entry-of-chunk)
//    pre_w0, pre_w1, pre_w2]       (weights entering the VLM pre-update == the
//                                   original w*; saved explicitly for symmetry)
std::vector<torch::Tensor> causal_ttt_forward_save(
    torch::Tensor w0, torch::Tensor w1, torch::Tensor w2,
    torch::Tensor q, torch::Tensor k, torch::Tensor v,
    torch::Tensor lr0, torch::Tensor lr1, torch::Tensor lr2,
    int64_t chunk_size,
    c10::optional<torch::Tensor> vlm_k,
    c10::optional<torch::Tensor> vlm_v,
    c10::optional<torch::Tensor> vlm_lr0,
    c10::optional<torch::Tensor> vlm_lr1,
    c10::optional<torch::Tensor> vlm_lr2) {

    TORCH_CHECK(q.is_cuda(), "causal_ttt_forward_save: inputs must be CUDA tensors");
    auto w0_norm = w0.detach().norm(2, 1, true);
    auto w1_norm = w1.detach().norm(2, 1, true);
    auto w2_norm = w2.detach().norm(2, 1, true);

    auto pre_w0 = w0, pre_w1 = w1, pre_w2 = w2;  // original weights (pre-update entry)
    if (vlm_k.has_value()) {
        fw_update(w0, w1, w2, vlm_k.value(), vlm_v.value(),
                  vlm_lr0.value(), vlm_lr1.value(), vlm_lr2.value(),
                  w0_norm, w1_norm, w2_norm);
    }

    const int64_t L = q.size(1);
    using torch::indexing::Slice;
    std::vector<torch::Tensor> outs, e0, e1, e2;
    for (int64_t s = 0; s < L; s += chunk_size) {
        int64_t e = std::min(s + chunk_size, L);
        // record entry weights of this chunk BEFORE its update
        e0.push_back(w0); e1.push_back(w1); e2.push_back(w2);
        auto qi = q.index({Slice(), Slice(s, e), Slice()});
        outs.push_back(fw_apply(qi, w0, w1, w2));
        auto ki = k.index({Slice(), Slice(s, e), Slice()});
        auto vi = v.index({Slice(), Slice(s, e), Slice()});
        auto l0 = lr0.index({Slice(), Slice(s, e), Slice()});
        auto l1 = lr1.index({Slice(), Slice(s, e), Slice()});
        auto l2 = lr2.index({Slice(), Slice(s, e), Slice()});
        fw_update(w0, w1, w2, ki, vi, l0, l1, l2, w0_norm, w1_norm, w2_norm);
    }
    auto output = torch::cat(outs, 1);
    auto entry_w0 = torch::stack(e0, 0);  // [n_chunk, B, d, dh]
    auto entry_w1 = torch::stack(e1, 0);
    auto entry_w2 = torch::stack(e2, 0);
    return {output, w0, w1, w2, entry_w0, entry_w1, entry_w2, pre_w0, pre_w1, pre_w2};
}

//
// Mirrors the proven torch manual backward (ttt_manual_backward.py):
// checkpoint-style BPTT. Forward pass recomputes + saves the ENTRY weights of
// each chunk (and the pre-update entry); reverse pass recomputes each chunk's
// intermediates from its entry weights and applies the vjp chain. GEMMs via
// ATen bmm; silu derivatives + the two normalize-vjps via custom kernels.

namespace {
using torch::indexing::Slice;
}  // namespace

// silu(x) helper via existing kernel (silu_glu(x, ones)=silu(x))
static inline torch::Tensor silu_only(const torch::Tensor& x) {
    return ttt_cuda::silu_glu(x, torch::ones_like(x));
}

std::vector<torch::Tensor> causal_ttt_backward(
    torch::Tensor w0, torch::Tensor w1, torch::Tensor w2,
    torch::Tensor q, torch::Tensor k, torch::Tensor v,
    torch::Tensor lr0, torch::Tensor lr1, torch::Tensor lr2,
    int64_t chunk_size,
    torch::Tensor g_out, torch::Tensor g_w0n, torch::Tensor g_w1n, torch::Tensor g_w2n,
    c10::optional<torch::Tensor> vlm_k,
    c10::optional<torch::Tensor> vlm_v,
    c10::optional<torch::Tensor> vlm_lr0,
    c10::optional<torch::Tensor> vlm_lr1,
    c10::optional<torch::Tensor> vlm_lr2,
    // Phase 1: precomputed per-chunk entry weights [n_chunk,B,*,*] from
    // causal_ttt_forward_save. When present, skip the forward-recompute loop.
    c10::optional<torch::Tensor> entry_w0 = c10::nullopt,
    c10::optional<torch::Tensor> entry_w1 = c10::nullopt,
    c10::optional<torch::Tensor> entry_w2 = c10::nullopt) {

    TORCH_CHECK(q.is_cuda(), "causal_ttt_backward: inputs must be CUDA");
    const double FEPS = 1e-7, WEPS = 1e-5;
    auto w0n_t = w0.norm(2, 1, true), w1n_t = w1.norm(2, 1, true), w2n_t = w2.norm(2, 1, true);
    bool has_vlm = vlm_k.has_value();
    bool have_entry = entry_w0.has_value();

    const int64_t L = q.size(1);
    std::vector<int64_t> starts;
    for (int64_t s = 0; s < L; s += chunk_size) starts.push_back(s);

    // ---- collect ENTRY weights of each chunk (and pre-vlm entry) ----
    std::vector<std::array<torch::Tensor, 3>> entry;  // weights entering each chunk
    std::array<torch::Tensor, 3> pre_entry = {w0, w1, w2};
    if (have_entry) {
        // Phase 1: use precomputed entry weights from forward_save -> NO recompute.
        auto ew0 = entry_w0.value(), ew1 = entry_w1.value(), ew2 = entry_w2.value();
        for (size_t ci = 0; ci < starts.size(); ++ci) {
            entry.push_back({ew0.select(0, ci), ew1.select(0, ci), ew2.select(0, ci)});
        }
        // pre_entry stays the original w* (== weights entering the vlm pre-update)
    } else {
        // fallback: recompute the forward to collect entry weights (old path)
        auto cw0 = w0, cw1 = w1, cw2 = w2;
        if (has_vlm) {
            fw_update(cw0, cw1, cw2, vlm_k.value(), vlm_v.value(),
                      vlm_lr0.value(), vlm_lr1.value(), vlm_lr2.value(), w0n_t, w1n_t, w2n_t);
        }
        for (size_t ci = 0; ci < starts.size(); ++ci) {
            int64_t s = starts[ci], e = std::min(s + chunk_size, L);
            entry.push_back({cw0, cw1, cw2});
            auto ki = k.index({Slice(), Slice(s, e), Slice()});
            auto vi = v.index({Slice(), Slice(s, e), Slice()});
            auto l0 = lr0.index({Slice(), Slice(s, e), Slice()});
            auto l1 = lr1.index({Slice(), Slice(s, e), Slice()});
            auto l2 = lr2.index({Slice(), Slice(s, e), Slice()});
            fw_update(cw0, cw1, cw2, ki, vi, l0, l1, l2, w0n_t, w1n_t, w2n_t);
        }
    }

    // ---- accumulators ----
    auto g_q = torch::zeros_like(q), g_k = torch::zeros_like(k), g_v = torch::zeros_like(v);
    auto g_lr0 = torch::zeros_like(lr0), g_lr1 = torch::zeros_like(lr1), g_lr2 = torch::zeros_like(lr2);
    auto gw0 = g_w0n.clone(), gw1 = g_w1n.clone(), gw2 = g_w2n.clone();

    // chunk-level vjp closure (also used for vlm pre-update with do_apply=false)
    auto chunk_vjp = [&](const torch::Tensor& W0, const torch::Tensor& W1, const torch::Tensor& W2,
                         const torch::Tensor& qi, const torch::Tensor& ki, const torch::Tensor& vi,
                         const torch::Tensor& l0, const torch::Tensor& l1, const torch::Tensor& l2,
                         const torch::Tensor& g_oi, bool do_apply,
                         torch::Tensor& out_gw0, torch::Tensor& out_gw1, torch::Tensor& out_gw2,
                         torch::Tensor& out_gq, torch::Tensor& out_gk, torch::Tensor& out_gv,
                         torch::Tensor& out_gl0, torch::Tensor& out_gl1, torch::Tensor& out_gl2) {
        // recompute forward intermediates
        auto gate = ki.bmm(W0), up = ki.bmm(W2);
        auto sd_g = ttt_cuda::silu_derivs(gate);  // silu'(gate), silu''(gate)
        auto sg = silu_only(gate);
        auto hidden = sg * up;
        auto dhidden = vi.bmm(W1.transpose(-1, -2));
        auto dhid_bm = dhidden * sg;
        auto dgate = dhidden * up;
        auto mm = sd_g[0];
        auto dgba = dgate * mm;
        // lr is fp32; A* promote to fp32 -> cast back to operand dtype before bmm
        // (matches the reference forward's `.to(vi.dtype())`).
        auto A0 = (ki * l0).to(dgba.dtype());
        auto A1 = (hidden * l1).to(vi.dtype());
        auto A2 = (ki * l2).to(dhid_bm.dtype());
        auto fn1_in = A1.transpose(-1, -2).bmm(vi);
        auto fn0_in = A0.transpose(-1, -2).bmm(dgba);
        auto fn2_in = A2.transpose(-1, -2).bmm(dhid_bm);
        auto rn1 = frob_normalize(fn1_in), rn0 = frob_normalize(fn0_in), rn2 = frob_normalize(fn2_in);
        auto w0_pre = W0 + rn0, w1_pre = W1 + rn1, w2_pre = W2 + rn2;

        // ---- vjp: weightnorm ----
        auto g_w0_pre = ttt_cuda::weightnorm_bwd(out_gw0, w0_pre, w0n_t, WEPS);
        auto g_w1_pre = ttt_cuda::weightnorm_bwd(out_gw1, w1_pre, w1n_t, WEPS);
        auto g_w2_pre = ttt_cuda::weightnorm_bwd(out_gw2, w2_pre, w2n_t, WEPS);
        // w_pre = w_old + raw
        auto g_w0_old = g_w0_pre.clone(), g_w1_old = g_w1_pre.clone(), g_w2_old = g_w2_pre.clone();
        auto g_rn0 = g_w0_pre, g_rn1 = g_w1_pre, g_rn2 = g_w2_pre;
        // frobnorm vjp
        auto g_fn0 = ttt_cuda::frobnorm_bwd(g_rn0, fn0_in, FEPS);
        auto g_fn1 = ttt_cuda::frobnorm_bwd(g_rn1, fn1_in, FEPS);
        auto g_fn2 = ttt_cuda::frobnorm_bwd(g_rn2, fn2_in, FEPS);

        // raw_w1 = A1^T @ vi
        auto g_A1 = vi.bmm(g_fn1.transpose(-1, -2));
        auto g_vi = A1.bmm(g_fn1);
        // lr is fp32; cast lr-scaled grads back to activation dtype for later bmm
        auto g_hidden = (g_A1 * l1).to(up.dtype());
        out_gl1.index({Slice(), Slice(), Slice()}) += (g_A1 * hidden).sum(-1, true);
        // raw_w0 = A0^T @ dgba
        auto g_A0 = dgba.bmm(g_fn0.transpose(-1, -2));
        auto g_dgba = A0.bmm(g_fn0);
        auto g_ki = (g_A0 * l0).to(ki.dtype());
        out_gl0.index({Slice(), Slice(), Slice()}) += (g_A0 * ki).sum(-1, true);
        // raw_w2 = A2^T @ dhid_bm
        auto g_A2 = dhid_bm.bmm(g_fn2.transpose(-1, -2));
        auto g_dhid_bm = A2.bmm(g_fn2);
        g_ki = g_ki + (g_A2 * l2).to(ki.dtype());
        out_gl2.index({Slice(), Slice(), Slice()}) += (g_A2 * ki).sum(-1, true);

        // dgba = dgate * m ; m=silu'(gate)
        auto g_dgate = g_dgba * mm;
        auto g_gate = g_dgba * dgate * sd_g[1];  // * silu''(gate)
        // dgate = dhidden*up
        auto g_dhidden = g_dgate * up;
        auto g_up = g_dgate * dhidden;
        // dhid_bm = dhidden*sg
        g_dhidden = g_dhidden + g_dhid_bm * sg;
        auto g_sg = g_dhid_bm * dhidden;
        // dhidden = vi @ w1^T
        g_vi = g_vi + g_dhidden.bmm(W1);
        g_w1_old = g_w1_old + g_dhidden.transpose(-1, -2).bmm(vi);
        // hidden = sg*up
        g_sg = g_sg + g_hidden * up;
        g_up = g_up + g_hidden * sg;
        // up = ki@w2
        g_ki = g_ki + g_up.bmm(W2.transpose(-1, -2));
        g_w2_old = g_w2_old + ki.transpose(-1, -2).bmm(g_up);
        // sg = silu(gate)
        g_gate = g_gate + g_sg * sd_g[0];
        // gate = ki@w0
        g_ki = g_ki + g_gate.bmm(W0.transpose(-1, -2));
        g_w0_old = g_w0_old + ki.transpose(-1, -2).bmm(g_gate);

        // accumulate k/v grads (cast to slice dtype: g_ki/g_vi may be fp32 due to lr)
        out_gk.index({Slice(), Slice(), Slice()}) += g_ki.to(out_gk.dtype());
        out_gv.index({Slice(), Slice(), Slice()}) += g_vi.to(out_gv.dtype());

        // ---- apply path ----
        if (do_apply) {
            auto gate_q = qi.bmm(W0), up_q = qi.bmm(W2);
            auto sq = silu_only(gate_q);
            auto h_q = sq * up_q;
            auto g_h_q = g_oi.bmm(W1.transpose(-1, -2));
            g_w1_old = g_w1_old + h_q.transpose(-1, -2).bmm(g_oi);
            auto g_sq = g_h_q * up_q;
            auto g_up_q = g_h_q * sq;
            auto g_qi = g_up_q.bmm(W2.transpose(-1, -2));
            g_w2_old = g_w2_old + qi.transpose(-1, -2).bmm(g_up_q);
            auto sd_q = ttt_cuda::silu_derivs(gate_q);
            auto g_gate_q = g_sq * sd_q[0];
            g_qi = g_qi + g_gate_q.bmm(W0.transpose(-1, -2));
            g_w0_old = g_w0_old + qi.transpose(-1, -2).bmm(g_gate_q);
            out_gq.index({Slice(), Slice(), Slice()}) += g_qi.to(out_gq.dtype());
        }

        out_gw0 = g_w0_old; out_gw1 = g_w1_old; out_gw2 = g_w2_old;
    };

    // ---- reverse over chunks ----
    for (int64_t idx = (int64_t)starts.size() - 1; idx >= 0; --idx) {
        int64_t s = starts[idx], e = std::min(s + chunk_size, L);
        auto W0 = entry[idx][0], W1 = entry[idx][1], W2 = entry[idx][2];
        auto qi = q.index({Slice(), Slice(s, e), Slice()});
        auto ki = k.index({Slice(), Slice(s, e), Slice()});
        auto vi = v.index({Slice(), Slice(s, e), Slice()});
        auto l0 = lr0.index({Slice(), Slice(s, e), Slice()});
        auto l1 = lr1.index({Slice(), Slice(s, e), Slice()});
        auto l2 = lr2.index({Slice(), Slice(s, e), Slice()});
        auto g_oi = g_out.index({Slice(), Slice(s, e), Slice()});
        auto gk_slice = g_k.index({Slice(), Slice(s, e), Slice()});
        auto gv_slice = g_v.index({Slice(), Slice(s, e), Slice()});
        auto gq_slice = g_q.index({Slice(), Slice(s, e), Slice()});
        auto gl0_slice = g_lr0.index({Slice(), Slice(s, e), Slice()});
        auto gl1_slice = g_lr1.index({Slice(), Slice(s, e), Slice()});
        auto gl2_slice = g_lr2.index({Slice(), Slice(s, e), Slice()});
        chunk_vjp(W0, W1, W2, qi, ki, vi, l0, l1, l2, g_oi, /*do_apply=*/true,
                  gw0, gw1, gw2, gq_slice, gk_slice, gv_slice, gl0_slice, gl1_slice, gl2_slice);
    }

    std::vector<torch::Tensor> result = {gw0, gw1, gw2, g_q, g_k, g_v, g_lr0, g_lr1, g_lr2};

    // ---- vlm pre-update (update-only) ----
    if (has_vlm) {
        auto dummy_q = torch::Tensor();
        auto g_vk = torch::zeros_like(vlm_k.value());
        auto g_vv = torch::zeros_like(vlm_v.value());
        auto g_vl0 = torch::zeros_like(vlm_lr0.value());
        auto g_vl1 = torch::zeros_like(vlm_lr1.value());
        auto g_vl2 = torch::zeros_like(vlm_lr2.value());
        auto g_oi_dummy = torch::Tensor();
        chunk_vjp(pre_entry[0], pre_entry[1], pre_entry[2], dummy_q,
                  vlm_k.value(), vlm_v.value(), vlm_lr0.value(), vlm_lr1.value(), vlm_lr2.value(),
                  g_oi_dummy, /*do_apply=*/false,
                  gw0, gw1, gw2, /*gq*/g_vk, g_vk, g_vv, g_vl0, g_vl1, g_vl2);
        // after pre-update vjp, gw0..gw2 are grads wrt the ORIGINAL w0/w1/w2
        result[0] = gw0; result[1] = gw1; result[2] = gw2;
        result.push_back(g_vk); result.push_back(g_vv);
        result.push_back(g_vl0); result.push_back(g_vl1); result.push_back(g_vl2);
    }
    return result;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("causal_ttt_forward", &causal_ttt_forward,
          "Causal block fast-weight SwiGLU TTT forward (CUDA)");
    m.def("causal_ttt_forward_save", &causal_ttt_forward_save,
          "Forward that also saves per-chunk entry weights (Phase 1 no-recompute backward)");
    m.def("causal_ttt_backward", &causal_ttt_backward,
          "Causal block fast-weight SwiGLU TTT backward (CUDA, Plan A)",
          py::arg("w0"), py::arg("w1"), py::arg("w2"),
          py::arg("q"), py::arg("k"), py::arg("v"),
          py::arg("lr0"), py::arg("lr1"), py::arg("lr2"),
          py::arg("chunk_size"),
          py::arg("g_out"), py::arg("g_w0n"), py::arg("g_w1n"), py::arg("g_w2n"),
          py::arg("vlm_k"), py::arg("vlm_v"),
          py::arg("vlm_lr0"), py::arg("vlm_lr1"), py::arg("vlm_lr2"),
          py::arg("entry_w0") = c10::nullopt,
          py::arg("entry_w1") = c10::nullopt,
          py::arg("entry_w2") = c10::nullopt);
    // expose backward primitives for unit testing
    m.def("silu_derivs", &ttt_cuda::silu_derivs, "silu' and silu''");
    m.def("frobnorm_bwd", &ttt_cuda::frobnorm_bwd, "Frobenius-normalize vjp");
    m.def("weightnorm_bwd", &ttt_cuda::weightnorm_bwd, "weight-norm (per-col) vjp");
    m.def("infer_step", &ttt_cuda::infer_step,
          "Fused single-token TTT inference apply + RMSNorm (CUDA)");
    m.def("infer_step_mid", &ttt_cuda::infer_step_mid,
          "Stage-1 mega: fused q-norm + apply + RMSNorm (CUDA)");
}