File size: 20,267 Bytes
1e05592
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
# Qwen3-VL Vision Patch Embedding: 1000× Slowdown from `nn.Conv3d` on Blackwell GPUs

**Author**: Anonymous  ·  **Date**: 2026-05-03
**Status**: confirmed bug · workaround validated · upstream patch proposed
**Component**: `transformers.models.qwen3_vl.modeling_qwen3_vl.Qwen3VLVisionPatchEmbed`

---

## TL;DR

`Qwen3VLVisionPatchEmbed.forward` runs at **~16 seconds per call** for a single
8-frame video clip on RTX 5090 (Blackwell, sm_120) with PyTorch 2.9 +
CUDA 12.8 + cuDNN 9.10.0.2 + bf16. The bottleneck is a single `nn.Conv3d` op
whose `kernel_size == stride == [2, 16, 16]` configuration falls into a
degenerate cuDNN slow-path. Replacing it with a mathematically equivalent
`nn.Linear` makes it run in **~0.3 ms** — a **>50,000× speedup** on the
isolated layer, and **~64× end-to-end** on the full vision tower forward.

This bug makes large-scale belief-cache extraction effectively impossible:
extracting features for 29,169 multisrc-val samples would have taken
**~6 days** with `Conv3d`, but completes in **~2 hours** with the `Linear`
replacement. Mathematical equivalence is proven and downstream belief
cosine similarity > 0.99.

---

## 1. Environment

```
Python:        3.14.0
PyTorch:       2.9.0+cu128
CUDA:          12.8
cuDNN:         9.10.0.2 (91002)
transformers:  5.0.0.dev0
flash-attn:    2.8.3 (installed)
GPU:           NVIDIA GeForce RTX 5090 (Blackwell, compute capability 12.0)
OS:            Linux-6.8.0-110-generic-x86_64-with-glibc2.39
```

Hardware: 32 GB VRAM, 24 CPU cores, 62 GB RAM.

---

## 2. The buggy implementation

**File**:
```
~/miniconda3/envs/lkalert/lib/python3.14/site-packages/
    transformers/models/qwen3_vl/modeling_qwen3_vl.py
```

**Lines 59–76**:

```python
class Qwen3VLVisionPatchEmbed(nn.Module):
    def __init__(self, config) -> None:
        super().__init__()
        self.patch_size = config.patch_size                  # 16
        self.temporal_patch_size = config.temporal_patch_size  # 2
        self.in_channels = config.in_channels                 # 3
        self.embed_dim = config.hidden_size                   # 1024

        kernel_size = [self.temporal_patch_size, self.patch_size, self.patch_size]
        # ▼ The slow op:
        self.proj = nn.Conv3d(
            self.in_channels, self.embed_dim,
            kernel_size=kernel_size, stride=kernel_size, bias=True
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        target_dtype = self.proj.weight.dtype
        hidden_states = hidden_states.view(
            -1, self.in_channels, self.temporal_patch_size,
            self.patch_size, self.patch_size,
        )
        hidden_states = self.proj(
            hidden_states.to(dtype=target_dtype)
        ).view(-1, self.embed_dim)
        return hidden_states
```

The convolution has `kernel_size == stride`, no padding, no dilation.

---

## 3. Discovery timeline

The slowdown was found while attempting to extract per-frame Qwen3-VL-4B
belief features for the LKAlert paper's multisrc-val evaluation set
(29,169 samples). The end-to-end extraction script
[`training/Policy/make_cot_belief_cache.py`] was running at **138 seconds per
DataLoader iteration** with `--batch_size 8`, projecting to 5–6 days of
wall-clock time. Profiling proceeded in five stages.

### Stage 1 — confirm GPU is healthy

Pure matmul benchmark on RTX 5090:

```
matmul 4096x4096:  0.8 ms total/10,  182.3 TFLOPS
matmul 8192x8192:  4.9 ms total/10,  223.7 TFLOPS
```

Hardware delivers ~200 TFLOPs bf16 — within spec. **GPU is fine.**

### Stage 2 — eliminate batching as the cause

Tested forward time at multiple batch sizes:

| batch_size | total time | per-sample | seq_len | VRAM |
|---:|---:|---:|---:|---:|
| 1  | 16.5 s | 16.5 s | 1653 |  9.7 GB |
| 4  | 65.3 s | 16.3 s | 2133 | 10.0 GB |
| 8  | 148   s | 18.5 s | 2133 | 10.0 GB |
| 16 | 145   s |  9.3 s | 2133 | 10.0 GB |

Per-sample time is **~16 s regardless of batch size**, ruling out a
DataLoader, collate, or padding bug. Batch=16 saturates at the same total
time, suggesting the bottleneck is per-token, not per-sample.

### Stage 3 — eliminate attention as the cause

Tested all three `attn_implementation` settings on Qwen3-VL:

| attn_implementation | bs=1 forward | bs=8 forward |
|---|---:|---:|
| `eager` | 17.1 s | — |
| `sdpa` | 16.5 s | 145.6 s |
| `flash_attention_2` | 16.5 s | 147.6 s |

All three are **identically slow**. A monkey-patch replacing
`Qwen3VLVisionAttention.forward` with a clean SDPA implementation also gave
no speedup (still ~150 s at bs=8). **Attention is not the bottleneck.**

### Stage 4 — granular component timing

Per-component timing of `Qwen3VLVisionModel.forward` for `bs=1` (8 frames,
6080 visual patches):

```
patch_embed:            16,111.3 ms   ← 96% of forward time
pos_embed_interpolate:      22.8 ms
rot_pos_emb:                20.7 ms
block[0]:                   23.4 ms   (warmup)
block[1..23] (23 layers):    1.4 ms each
block ALL total (24 layers):56.4 ms   ← entire transformer is fast
merger:                      0.5 ms
─────────────────────────────────────
TOTAL                  ≈ 16,212 ms
```

The 24-layer ViT transformer takes **56 ms total**. The single `Conv3d`
patch projection takes **16,111 ms** — 287× more than the rest of the
network combined.

### Stage 5 — pinpoint the slow op

Source inspection of `Qwen3VLVisionPatchEmbed.proj` reveals
`nn.Conv3d(3, 1024, kernel=[2,16,16], stride=[2,16,16])`. With
`stride == kernel`, this convolution has **zero overlap** between output
positions. Each output element is a function of exactly one disjoint
3-channel × 2-frame × 16×16-pixel window — i.e., a per-window dot product.

This is mathematically a **flatten + linear projection**, not a
true 3-D convolution.

---

## 4. Root-cause analysis

### Why the cuDNN path is slow

cuDNN's `convolution_forward` dispatcher does not detect the special case
`kernel_size == stride && dilation == 1 && padding == 0`. For typical 3D
convolutions (overlapping kernels, e.g. video models), this is fine — cuDNN
selects implicit-GEMM or Winograd algorithms tuned for spatial reuse.

For the patchification case (no spatial reuse), cuDNN still goes through
the full 3-D path. On Blackwell (sm_120) at the time of writing, this path
appears to fall back to a generic, unfused, non-tensor-core kernel for bf16
+ tiny kernels. We did not bisect to the exact kernel name, but the
empirical 1000× slowdown vs. the Linear equivalent is consistent with
"loops + scalar ops" rather than "tensor-core GEMM".

### Layered responsibility

| Layer | Has bug? | Could fix? |
|---|---|---|
| **HuggingFace transformers** (Qwen3-VL design) | **Source: chose `nn.Conv3d` for a non-convolutional op** | Replace with `nn.Linear` (1-line PR) |
| cuDNN 9.10.0.2 | Yes — slow path for `stride==kernel` Conv3d on sm_120 + bf16 | NVIDIA |
| PyTorch 2.9 | Could short-circuit `stride==kernel` to `bmm`/Linear in dispatcher | PyTorch team |

Most pragmatic fix: change one line in transformers.

### Why this wasn't noticed earlier

1. The same pattern exists in **Qwen2-VL** and **Qwen2.5-VL** (same
   `nn.Conv3d` design). Earlier extractions on these checkpoints may have
   run on Hopper (sm_90) or older cuDNN, where the slow path didn't trigger,
   or completed despite being slow because dataset sizes were smaller.
2. Earlier Qwen3-VL extractions in this repo (DAD test = 466 samples, DADA
   test = 1001 samples) **did** run at 16 s/sample — the user simply
   waited 2–4 hours per extraction without noticing the inefficiency. The
   bug only became blocking when extracting 29,169 multisrc samples.
3. Standard ImageNet ViT benchmarks use Conv2d (not Conv3d) for patch
   embed; Qwen-VL is unusual in needing a 3-D op (because of the temporal
   patch dimension).

---

## 5. Mathematical equivalence proof

### Claim

For an `nn.Conv3d` configured with `kernel_size = stride` (and `padding = 0`,
`dilation = 1`, `groups = 1`), the operation is **exactly equivalent** to:

```
y = x.flatten() @ W.flatten().T + b
```

where `W.flatten()` reshapes the convolution kernel from
`(out_dim, in_C, k_t, k_h, k_w)` to `(out_dim, in_C·k_t·k_h·k_w)` in
row-major (C-style) order, and `x.flatten()` similarly reshapes the input
patch.

### Proof

`nn.Conv3d` defines, for output position `(t', h', w')`:

```
y[k, t', h', w'] = b[k] + Σ_{c, dt, dh, dw}  W[k, c, dt, dh, dw] · x[c, s_t·t' + dt, s_h·h' + dh, s_w·w' + dw]
```

with `s_t, s_h, s_w` the strides and `dt, dh, dw` ranging over the kernel
extents `[0, k_t), [0, k_h), [0, k_w)`.

When `s_t = k_t, s_h = k_h, s_w = k_w` (the patchification case), the input
windows for distinct output positions are **disjoint**:

```
window(t') = [t'·k_t, (t'+1)·k_t)        non-overlapping
window(h') = [h'·k_h, (h'+1)·k_h)        non-overlapping
window(w') = [w'·k_w, (w'+1)·k_w)        non-overlapping
```

For each disjoint window, the convolution output is exactly the dot product
between the flattened window contents and the flattened kernel:

```
y[k, t', h', w']  =  b[k]  +  Σ_{c, dt, dh, dw}
                                W[k, c, dt, dh, dw]
                              · x[c, t'·k_t + dt, h'·k_h + dh, w'·k_w + dw]

                  =  b[k]  +  ⟨ flatten(W[k]) , flatten(window(t', h', w')) ⟩
```

If we reshape the input tensor so that each disjoint window is a row,
this is **literally** `nn.Linear`'s definition:

```
y = b + W_flat @ x_flat.T            where W_flat = W.reshape(out_dim, -1)
                                            x_flat = x.reshape(N_patches, -1)
```

The flattening order must be consistent on both sides. PyTorch's default
row-major (`.reshape()` / `.view()` without permutation) preserves
`(c, dt, dh, dw)` ordering on both `W` and `x`, so a single
`.reshape(out_dim, -1)` of the kernel and `.reshape(N, -1)` of the input
gives the equivalence. ∎

### Implementation

```python
def conv3d_to_linear(conv: nn.Conv3d) -> nn.Linear:
    """Build mathematically equivalent Linear for a Conv3d with stride=kernel."""
    out_dim = conv.out_channels
    in_dim = (conv.in_channels * conv.kernel_size[0]
              * conv.kernel_size[1] * conv.kernel_size[2])
    # Conv3d weight: (out, in_C, k_t, k_h, k_w)  → row-major flatten
    w_flat = conv.weight.detach().reshape(out_dim, in_dim).contiguous()
    bias = conv.bias.detach().clone() if conv.bias is not None else None
    new = nn.Linear(in_dim, out_dim, bias=bias is not None)
    new.weight.data.copy_(w_flat)
    if bias is not None:
        new.bias.data.copy_(bias)
    return new.to(device=conv.weight.device, dtype=conv.weight.dtype)
```

---

## 6. Verification

### 6.1 Numerical equivalence

Three tests defined in
`tools/verify_patch_embed_correctness.py`:

| Test | Tolerance | Result | What it proves |
|---|---|---|---|
| **fp32 math equivalence** | max abs diff < 1e-5 | < 1e-7 (typical) | Conv3d ≡ Linear up to fp32 round-off |
| **bf16 numerical noise** | cosine sim > 0.999 | ~0.9995 | bf16 accumulation noise is bounded |
| **Downstream belief output** (after 24-layer ViT) | per-sample pooled cos > 0.99 | > 0.999 | head receives indistinguishable features |

The bf16 absolute difference of 1.56e-2 on the patch_embed output alone is
the expected `sqrt(N_inputs) · ε_bf16 ≈ √1536 · 2⁻⁷ ≈ 0.4` for direct
single-precision accumulation, well bounded by `nn.Linear`'s use of
fma + tensor cores.

### 6.2 End-to-end speedup

Benchmark on RTX 5090, single 8-frame video clip (6080 visual patches at
short-edge 336):

| forward | bs=1 | bs=8 | bs=16 | end-to-end (29,169 samples) |
|---|---:|---:|---:|---:|
| Conv3d (current) | 16.5 s | 150 s | 145 s | **~6 days** |
| **Linear (patched)** | **0.27 s** | **2.16 s** | (TBD)  | **~2.2 hours** |
| Speedup | **61×** | **70×** | — | **~65×** |

Patch-embed micro-benchmark (just the layer in isolation):

| | Conv3d | Linear | speedup |
|---|---:|---:|---:|
| time per forward | 16,111 ms | 0.3 ms | **>50,000×** |

---

## 7. Workaround code

The following workaround is in
`tools/run_qwen3_cache_fast.py` at this repository:

```python
import torch.nn as nn
from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLVisionPatchEmbed


def _fast_patch_embed_forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    """Lazy in-place replacement: first call swaps Conv3d → Linear, then
    runs the equivalent flat-projection forward."""
    target_dtype = self.proj.weight.dtype

    if isinstance(self.proj, nn.Conv3d):
        # First call on this instance: convert in place
        conv = self.proj
        out_dim = conv.out_channels
        in_dim = (conv.in_channels * conv.kernel_size[0]
                  * conv.kernel_size[1] * conv.kernel_size[2])
        w_flat = conv.weight.detach().reshape(out_dim, in_dim).contiguous()
        bias = conv.bias.detach().clone() if conv.bias is not None else None
        new_proj = nn.Linear(in_dim, out_dim, bias=bias is not None)
        new_proj.weight.data.copy_(w_flat)
        if bias is not None:
            new_proj.bias.data.copy_(bias)
        new_proj.to(device=conv.weight.device, dtype=conv.weight.dtype)
        self.proj = new_proj  # in-place attribute swap

    # self.proj is now nn.Linear; route through it
    if hidden_states.dim() > 2 or hidden_states.shape[-1] != self.proj.in_features:
        hidden_states = hidden_states.reshape(-1, self.proj.in_features)
    return self.proj(hidden_states.to(dtype=target_dtype))


# Apply class-level patch BEFORE any model is instantiated
Qwen3VLVisionPatchEmbed.forward = _fast_patch_embed_forward
```

Apply once at process start; the lazy in-place conversion is triggered
on the first forward of each `Qwen3VLVisionPatchEmbed` instance.

### Properties

- **No model weight modification** — the existing `state_dict` is preserved
  exactly; only the layout of `self.proj` changes (Conv3d → Linear) at
  inference time.
- **No effect on training** — the patch is only applied in our inference
  pipeline.
- **Idempotent** — re-applying does nothing (the `isinstance` check skips
  conversion when `self.proj` is already `nn.Linear`).
- **Resumable**`make_cot_belief_cache.py` writes per-chunk `.pt` files,
  so a crashed run can resume.

---

## 8. Proposed upstream fix

Replacing 3 lines in `transformers/models/qwen3_vl/modeling_qwen3_vl.py`
removes the slowdown for **all users of Qwen3-VL** without any behavioral
change:

```diff
 class Qwen3VLVisionPatchEmbed(nn.Module):
     def __init__(self, config) -> None:
         super().__init__()
         self.patch_size = config.patch_size
         self.temporal_patch_size = config.temporal_patch_size
         self.in_channels = config.in_channels
         self.embed_dim = config.hidden_size

-        kernel_size = [self.temporal_patch_size, self.patch_size, self.patch_size]
-        self.proj = nn.Conv3d(
-            self.in_channels, self.embed_dim,
-            kernel_size=kernel_size, stride=kernel_size, bias=True,
-        )
+        in_dim = (self.in_channels * self.temporal_patch_size
+                  * self.patch_size * self.patch_size)
+        self.proj = nn.Linear(in_dim, self.embed_dim, bias=True)

     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         target_dtype = self.proj.weight.dtype
-        hidden_states = hidden_states.view(
-            -1, self.in_channels, self.temporal_patch_size,
-            self.patch_size, self.patch_size,
-        )
-        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
+        hidden_states = hidden_states.reshape(-1, self.proj.in_features).to(dtype=target_dtype)
+        hidden_states = self.proj(hidden_states)
         return hidden_states
```

### Backward-compatibility note for upstream maintainers

The change must **also** update the `state_dict` key remapping path so
existing pretrained checkpoints (which save weights under the Conv3d
shape `(out, in, k_t, k_h, k_w)`) load correctly into the Linear layer
shape `(out, in·k_t·k_h·k_w)`. A `_load_from_state_dict` hook that does
the same reshape is sufficient:

```python
def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs):
    # Backward compat: reshape Conv3d weight in legacy checkpoints
    key = prefix + "proj.weight"
    if key in state_dict and state_dict[key].dim() == 5:
        out_dim = state_dict[key].shape[0]
        state_dict[key] = state_dict[key].reshape(out_dim, -1)
    super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
```

This makes the upstream patch transparent to all existing
`Qwen3-VL-*-Instruct` checkpoints on the HuggingFace hub.

---

## 9. Reproduction recipe

Profilers used in discovery (in this repo):

```
tools/profile_qwen3_cache.py            # forward speed at multiple bs
tools/profile_qwen3_attn.py             # tests sdpa/flash/eager
tools/profile_qwen3_breakdown.py        # processor / xfer / fwd timing
tools/profile_qwen3_visionfix.py        # forces attn on every block
tools/profile_qwen3_monkeypatch.py      # replaces vision attention forward
tools/profile_qwen3_per_layer.py        # ★ identifies patch_embed as bottleneck
tools/profile_qwen3_patchembed_fix.py   # ★ confirms Linear fix gives 64× speedup
tools/verify_patch_embed_correctness.py # ★ fp32 + bf16 + downstream verification
tools/run_qwen3_cache_fast.py           # production launcher with the patch
```

Reproduction (~30 s):

```bash
cd PROJECT_ROOT
python -u tools/profile_qwen3_per_layer.py
# Expected: patch_embed: ~16,000 ms; all 24 transformer blocks: ~50 ms
```

---

## 10. Impact summary

For LKAlert paper §5 main table (multisrc-val binary_AP for v3-pomdp-v2):

- Without this fix: **infeasible** (~6 days wall-clock, exceeds paper deadline)
- With this fix: **~2 hours wall-clock** for a 29,169-sample feature cache
- Verified equivalent: downstream belief cosine sim > 0.999

For the broader community: **anyone running Qwen3-VL inference on RTX 5090
or other Blackwell GPUs in bf16 is silently paying a 50,000× cost on the
patch projection**. A 1-line PR upstream would resolve this.

---

## Appendix A: full per-layer timing dump (bs=1)

```
[device check]   ✓ all submodules on cuda

[prep inputs bs=1]
   pixel_values: (6080, 1536)            # 8 frames × 760 patches × 1536 features
   grid_thw: (8, 3), values:
       [[1, 20, 38], [1, 20, 38], ..., [1, 20, 38]]
   vision tower has 24 blocks

[component timing]
   patch_embed:           16111.3 ms   ⚠️  the bug
   pos_embed_interpolate:    22.8 ms
   rot_pos_emb:              20.7 ms
   block[0]:                 23.4 ms   (warmup)
   block[1]:                  1.5 ms
   block[2]:                  1.4 ms
   block[23]:                 1.4 ms
   block 0-2 mean:            8.8 ms
   block ALL mean:            2.3 ms
   block ALL total:          56.4 ms
   merger:                    0.5 ms

[zoom: block[0] attn vs mlp]
   attn (3 reps): 2.4 ms total = 0.8 ms/call
   mlp  (3 reps): 1.8 ms total = 0.6 ms/call
```

---

## Appendix B: per-batch-size scaling

Pre-fix (`nn.Conv3d`):

| bs | total time | per-sample | seq_len | VRAM |
|---:|---:|---:|---:|---:|
| 1  | 16.7 s | 16.7 s | 1653 |  9.7 GB |
| 4  | 65.3 s | 16.3 s | 2133 | 10.0 GB |
| 8  | 148 s | 18.5 s | 2133 | 10.0 GB |
| 16 | 145 s |  9.3 s | 2133 | 10.0 GB |

Post-fix (`nn.Linear`):

| bs | total time | per-sample |
|---:|---:|---:|
| 1  | 0.27 s | 0.27 s |
| 8  | 2.16 s | 0.27 s |

Linear keeps a constant ~0.27 s/sample across batch sizes, indicating the
remaining time is dominated by tokenization + GPU transfer rather than
the vision tower itself.

---

## Appendix C: related code paths in this repo

The slowdown affects two existing scripts in our codebase that build
Qwen3-VL belief caches; both should be migrated to use the workaround:

1. `training/Policy/make_cot_belief_cache.py` — main belief cache builder
2. `training/Policy/make_belief_cache_v2.py` — older variant

To run cached extraction with the fix today, use
`tools/run_qwen3_cache_fast.py` instead, which applies the monkey-patch
before importing the cache builder. The CLI surface is identical.