Qwen 3 vl #359

loupicaaa · 2025-11-25T15:15:33Z

Qwen3VL bf16 example implementation based on the implementation of transformers for Qwen3VL 4B. Compile with a known seq length and a known resize shape for the image.
Support BMP (own implementation), jpg et png by using zignal library (fork) (added in third_party).

… main (adding bfactory fork library)

gwenzek · 2025-11-25T21:08:34Z

examples/qwen_3_vl/main.zig

+    };
+
+    // Initialize ZML platform
+    const create_opts_json = cli.args.@"create-options" orelse "{\"cpu\": {\"device_count\": 1}}";


dude, you broke github

(this was a joke)

but I don't think that device count 1 should be the default.
Also json is not really nice to type on CLI, it's better if you can switch to zon.

gwenzek · 2025-11-26T08:18:05Z

zml/exe.zig

-        //         ev.await(self.platform.pjrt_api) catch unreachable;
-        //     }
-        // }
+        for (events[0..sharding.num_partitions]) |e| {


We prefer to keep call asynchronous, you may need to add a few await() in your main to make it work like you want.

zml/hostbuffer.zig

zml/nn.zig

gwenzek · 2025-11-26T10:19:36Z

examples/qwen3_vl/main.zig

+        std.ascii.toLower(ext[2]) == 'p');
+
+    var height: u32 = undefined;
+    var width: u32 = undefined;


avoid this pattern.

you can do:

const image: RgbImage = if (is_bmp) img: { ... break :img img; } else img: { ... break :img img; };

gwenzek · 2025-11-26T10:23:30Z

examples/qwen_3_vl/main.zig

+    }
+}
+
+pub fn preprocessor(allocator: std.mem.Allocator, tokenizer: zml.tokenizer.Tokenizer, prompt: []const u8, config: qwen.Qwen.Config, preprocessor_config: PreprocessorConfig, image_path: []const u8, max_seq_len: u32, max_side: u32) !struct { image_buffer_chw: zml.HostBuffer, prompt_tokens: zml.HostBuffer, prompt_shape: zml.HostBuffer, image_dim: zml.HostBuffer, token_index: zml.HostBuffer, h_resized: u32, w_resized: u32 } {


can you give a name to this struct ? eg Input ? and put the deinit inside

gwenzek · 2025-11-26T10:26:19Z

examples/qwen_3_vl/qwen_3_vl.zig

+        const merge_size = self.qwen.config.vision_config.spatial_merge_size;
+
+        // Reshape the image to the correct shape for the vision transformer
+        image_chw_rescaled_normalized = image_chw_rescaled_normalized.reshape(.{ .temporal_patch_size = temporal_patch_size, .c = 3, .h_div = @divExact(grid_h, merge_size), .m1 = merge_size, .patch1 = patch_size, .w_div = @divExact(grid_w, merge_size), .m2 = merge_size, .patch2 = patch_size });


in general reshape( divisions ) is better served by a .splitAxis: no room for errors

gwenzek · 2025-11-26T10:26:32Z

examples/qwen_3_vl/qwen_3_vl.zig

+
+        // Reshape the image to the correct shape for the vision transformer
+        image_chw_rescaled_normalized = image_chw_rescaled_normalized.reshape(.{ .temporal_patch_size = temporal_patch_size, .c = 3, .h_div = @divExact(grid_h, merge_size), .m1 = merge_size, .patch1 = patch_size, .w_div = @divExact(grid_w, merge_size), .m2 = merge_size, .patch2 = patch_size });
+        image_chw_rescaled_normalized = image_chw_rescaled_normalized.transpose(.{ 2, 5, 3, 6, 1, 0, 4, 7 });


please use the names rather than the offsets

gwenzek · 2025-11-26T10:27:55Z

zml/tensor.zig


        if (args.steps != 1) {
-            res = res.scale(args.steps);
+            res = res.scale(range / @as(f64, @floatFromInt(args.steps - 1)));


thanks for the fix

gwenzek · 2025-11-26T10:28:25Z

Thanks for adding this !

examples/qwen3_vl/BUILD.bazel

gwenzek · 2025-11-26T14:06:57Z

examples/qwen_3_vl/qwen_3_vl.zig

+        const w = @divExact(image_grid_thw[2], 2);
+
+        // Repeat the index along the 3 dimensions based on grid size (after the text (+4 tokens according to the chat template))
+        const t_index = zml.Tensor.iota(Shape.init(.{ .t = t }, .i32), .t).reshape(.{ .t = -1, .hw = 1 }).repeat1d(1, h * w).flatten().addConstant(4);


This is what I have:

// Build the 3D positional ids for the vision transformer fn buildVisionPositionIds(position_ids: Tensor, image_grid_thw: [3]u32) struct { temporal: Tensor, height: Tensor, width: Tensor } { const t = image_grid_thw[0]; const h = @divExact(image_grid_thw[1], 2); const w = @divExact(image_grid_thw[2], 2); const img_position = zml.Tensor.scalar(4, .i32); // Repeat the index along the 3 dimensions based on grid size (after the text (+4 tokens according to the chat template)) const t_index = zml.Tensor.iota(.init(.{ .t = t, .h = h, .w = w }, .i32), .t).add(img_position).merge(.{ .seq = .{ .t, .h, .w }}); const h_index = zml.Tensor.iota(.init(.{ .h = h, .w = w }, .i32), .h).add(img_position).merge(.{ .seq = .{ .h, .w }}); const w_index = zml.Tensor.iota(.init(.{ .h = h, .w = w }, .i32), .w).add(img_position).merge(.{ .seq = .{ .h, .w }}); // Update the position ids with the 3D positional ids const temporal_ids = position_ids.scatterSlices(.{ .seq = img_position}, t_index, .{}); const height_ids = position_ids.scatterSlices(.{ .seq = img_position}, h_index, .{}); const width_ids = position_ids.scatterSlices(.{ .seq = img_position}, w_index, .{}); // For each dim, the output is the position ids (from 4 to grid size / merge size) repeated to reach the grid size return .{ .temporal = temporal_ids, .height = height_ids, .width = width_ids, }; }

But note that something is fishy: h_index and w_index don't have the same shape than t_index. IIUC the python code, t_index has the correct shape.

By using scatterSlices instead of dynamicUpdateSlice it works with/without the batch axis, so no need for unsqueeze.

gwenzek · 2025-11-26T14:25:55Z

examples/qwen_3_vl/qwen_3_vl.zig

+    // Build the 3D positional ids for the vision transformer
+    fn buildVisionPositionIds(position_ids: Tensor, image_grid_thw: [3]u32) struct { temporal: Tensor, height: Tensor, width: Tensor } {
+        const t = image_grid_thw[0];
+        const h = @divExact(image_grid_thw[1], 2);


why div by 2 ?

It's div by the spatial merge size (I hardcoded it I fix it). It is the side length of the square block of vision patches merged into a single visual token (that's why the number of token is timeheightwidth/spatial_merge_size^2). So for each side grid_h, grid_w is div to match the number of token image

gwenzek · 2025-11-26T14:30:34Z

examples/qwen_3_vl/qwen_3_vl.zig

+        const w = @divExact(image_grid_thw[2], 2);
+
+        // Repeat the index along the 3 dimensions based on grid size (after the text (+4 tokens according to the chat template))
+        const t_index = zml.Tensor.iota(Shape.init(.{ .t = t }, .i32), .t).reshape(.{ .t = -1, .hw = 1 }).repeat1d(1, h * w).flatten().addConstant(4);


But even beyond that I don't think the code is correct. Here is the output I have for temporal_ids:

info(zml): Device buffer: Buffer({1,512,i32})@738100e800: { {0,1,2,3,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166,167,168,169,170,171,172,173,174,175,176,177,178,179,180,181,182,183,184,185,186,187,188,189,190,191,192,193,194,195,196,197,198,199,200,201,202,203,204,205,206,207,208,209,210,211,212,213,214,215,216,217,218,219,220,221,222,223,224,225,226,227,228,229,230,231,232,233,234,235,236,237,238,239,240,241,242,243,244,245,246,247,248,249,250,251,252,253,254,255,256,257,258,259,260,261,262,263,264,265,266,267,268,269,270,271,272,273,274,275,276,277,278,279,280,281,282,283,284,285,286,287,288,289,290,291,292,293,294,295,296,297,298,299,300,301,302,303,304,305,306,307,308,309,310,311,312,313,314,315,316,317,318,319,320,321,322,323,324,325,326,327,328,329,330,331,332,333,334,335,336,337,338,339,340,341,342,343,344,345,346,347,348,349,350,351,352,353,354,355,356,357,358,359,360,361,362,363,364,365,366,367,368,369,370,371,372,373,374,375,376,377,378,379,380,381,382,383,384,385,386,387,388,389,390,391,392,393,394,395,396,397,398,399,400,401,402,403,404,405,406,407,408} },

I think it should read .{ 0, 1, 2, 3, 4, 4, 4, ..., 4, 4, 5, 6, 7, ... } but instead we have this cliff: 4, 4, 4, 21.

Edit:

we checked together the reference implem and it is indeed correct.

…s an output

examples/qwen3_vl/qwen3_vl.zig

gwenzek · 2025-12-01T15:46:06Z

examples/qwen3_vl/qwen3_vl.zig

+        position_ids = position_ids.dynamicUpdateSlice(.{ .seq = zml.Tensor.scalar(0, .i32) }, before_image_positions);
+
+        // Repeat the index along the 3 dimensions based on grid size (after the text (+4 tokens according to the chat template))
+        const t_index = zml.Tensor.iota(.init(.{ .bs = input_ids.dim(.bs), .t = t, .hw = h * w }, .i32), .t).reshape(.{ .bs = input_ids.dim(.bs), .seq = -1 }).add(text_before_image);


use an intermediate variables to avoid repetition. t_index, h_idnex and w_index can all use the same shape for the iota and the reshape.

examples/qwen3_vl/qwen3_vl.zig

gwenzek · 2025-12-01T16:19:07Z

examples/qwen3_vl/qwen3_vl.zig

+        return .{ stacked_position_ids, mrope_position_deltas };
+    }
+
+    test "buildVisionPositionIds" {


this is way too complicated for a simple test, here is a refactoring (assuming text_with_image is replaced by an integer)

test "buildVisionPositionIds" { std.debug.print("buildVisionPositionIds test started\n", .{}); const platform = zml.testing.env(); const allocator = std.testing.allocator; // Parameters const batch_size: u32 = 1; const seq_len: u32 = 79; // Create input buffers var input_ids_data = try allocator.alloc(i32, batch_size * seq_len); defer allocator.free(input_ids_data); for (0..batch_size * seq_len) |i| { input_ids_data[i] = @intCast(i % seq_len); } const input_ids_d = try zml.Buffer.fromSlice(platform, .{ .bs = batch_size, .seq = seq_len }, input_ids_data); defer input_ids_d.deinit(); const prompt_shape_d = try zml.Buffer.fromSlice(platform, .{ .seq = 3 }, &[_]i32{ 4, 64, 11 }); defer prompt_shape_d.deinit(); // Compile and execute buildVisionPositionIds const Local = struct { pub fn positionIds(input_ids: zml.Tensor, prompt_shape: zml.Tensor) zml.Tensor { return buildVisionPositionIds(2, input_ids, seq_len, prompt_shape, .{ 1, 16, 16 })[0]; } }; const result = try zml.testing.compileAndCall( platform, Local.positionIds, .{ input_ids_d, prompt_shape_d }, ); defer result.deinit(); const expected = [3][79]i32{ // temporal .{ 0, 1, 2, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22 }, // height .{ 0, 1, 2, 3, 4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10, 10, 10, 10, 10, 11, 11, 11, 11, 11, 11, 11, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22 }, // width .{ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 4, 5, 6, 7, 8, 9, 10, 11, 4, 5, 6, 7, 8, 9, 10, 11, 4, 5, 6, 7, 8, 9, 10, 11, 4, 5, 6, 7, 8, 9, 10, 11, 4, 5, 6, 7, 8, 9, 10, 11, 4, 5, 6, 7, 8, 9, 10, 11, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22 }, }; try std.testing.expectEqual(expected, try result.getValue([3][79]i32)); }

gwenzek · 2025-12-01T16:20:41Z

examples/qwen3_vl/qwen3_vl.zig

+}
+
+pub fn applyRotaryPositionalEmbedding(q: Tensor, k: Tensor, cos: Tensor, sin: Tensor) struct { Tensor, Tensor } {
+    const cos_q_unsqueezed = zml.torch.unsqueeze(cos, -2).broad(q.shape());


why do you need unsqueeze before broadcasting ? broadcasting should insert the axis if needed. See zml.nn.rope

gwenzek · 2025-12-01T16:24:31Z

examples/qwen3_vl/qwen3_vl.zig

+
+        // Interleaved mrope
+        // Slice the frequency tensor to get the frequency for the temporal, height and width dimensions
+        var freqs_t = freqs.slice1d(0, .{ .start = 0, .end = 1 }).squeeze(0).withTags(.{ .bs, .seq, .dh });


gwenzek · 2025-12-01T16:25:51Z

examples/qwen3_vl/qwen3_vl.zig

+        // Repeat 3 times (t, h, w) on dim 0, pos id 1dim on dim 1 (number of images I think) (3, batch size, dim_head//2, 1)
+        const inv_freq_expanded = inv_freq.reshape(.{ -1, 1 });
+        const position_ids_expanded = position_ids.reshape(.{ 3, @as(u32, @intCast(position_ids.dim(1))), 1, -1 }).convert(.f32); // (3, bs, 1, seq len)
+        var freqs = inv_freq_expanded.matmul(position_ids_expanded).transpose(.{ 0, 1, 3, 2 }); // (3, bs, dim_head//2, seq len)


please no matmul use, dot instead and specify the axis you want to contract over. This should avoid the reshapes/transpose.

gwenzek · 2025-12-01T16:26:40Z

examples/qwen3_vl/qwen3_vl.zig

+        // Gather scatter the frequencies to build the tensor such as [t,h,w,t,h,w,...,t,h,w,t,t,t,t]
+        const h_input = freqs_h.gather(.{ .dh = h_indices }, .{ .indices_are_sorted = true });
+        const w_input = freqs_w.gather(.{ .dh = w_indices }, .{ .indices_are_sorted = true });
+        freqs_t = freqs_t.transpose(.{ 2, 0, 1 });


avoid axis as integer, use the names

gwenzek · 2025-12-01T16:27:31Z

zml/hostbuffer.zig

-            const sliced_self = self.slice1d(0, .{ .start = di, .end = di + 1 }).squeeze(0);
-            try sliced_self.prettyPrintIndented(writer, num_rows, indent_level + 2, options);
-        }
+        // if (n < num_rows) return;


gwenzek · 2025-12-01T16:29:52Z

zml/testing.zig

    log.info("all good for {s} !", .{name});
 }

+pub fn testLayerWithoutInput(


I would expect testLayer to work with an empty argument list. Is there a bug there ? can we fix that instead of introducing a small variant ?

gwenzek · 2025-12-04T08:44:27Z

examples/qwen3_vl/qwen3_vl.zig

+        image_grid_thw: [3]u32,
+    ) struct { Tensor, Tensor } {
+        // Get the number of text tokens before the image, the number of image tokens and the number of text tokens after the image
+        const text_before_image = prompt_shape.choose1d(0, 0).convert(.i32);


the content of prompt_shape should be documented.
And I guess this should also be a chunkExact(3)

gwenzek · 2025-12-04T08:46:32Z

examples/qwen3_vl/qwen3_vl.zig

+        const t = grid_thw[0];
+        const h = grid_thw[1];
+        const w = grid_thw[2];
+        var hpos_ids = zml.torch.unsqueeze(zml.Tensor.arange(.{ .start = 0, .end = h, .step = 1 }, .f32), 1).repeat1d(1, @as(u63, @intCast(w)));


gwenzek · 2025-12-04T08:46:41Z

examples/qwen3_vl/qwen3_vl.zig

+        const h = grid_thw[1];
+        const w = grid_thw[2];
+        var hpos_ids = zml.torch.unsqueeze(zml.Tensor.arange(.{ .start = 0, .end = h, .step = 1 }, .f32), 1).repeat1d(1, @as(u63, @intCast(w)));
+        hpos_ids = hpos_ids.reshape(.{ .h_div = @divExact(h, m_size), .m1 = m_size, .w_div = @divExact(w, m_size), .m2 = m_size });


gwenzek · 2025-12-04T08:58:47Z

examples/qwen3_vl/qwen3_vl.zig

+
+pub const VisionRotaryEmbedding = struct {
+    rope_opts: zml.nn.RopeOpts,
+    dim: u32,


what is the dim used for ?

gwenzek · 2025-12-04T09:00:22Z

examples/qwen3_vl/qwen3_vl.zig

+    rope_opts: zml.nn.RopeOpts,
+    dim: u32,
+
+    pub fn init(allocator: std.mem.Allocator, dim: u32, theta: f32) !VisionRotaryEmbedding {


why allocatoor ?

my feeling is that this struct should not exist, put the options directly into the parent struct,
and the forward can be inlined or maybe a helper function.

gwenzek · 2025-12-04T09:26:29Z

examples/qwen3_vl/main.zig

+    const prompt_buffer = try allocator.alloc(u32, max_seq_len);
+
+    // Create the HostBuffers for the prompt, prompt shape, image size and token index
+    @memcpy(prompt_buffer[0..prompt_encoded.len], prompt_encoded);


why copy here ? you could reuse the already allocated memory ?

gwenzek · 2025-12-04T09:27:03Z

examples/qwen3_vl/main.zig

+    const number_image_pad_tokens = 1 * (@as(u32, @intFromFloat(h_resized)) / patch_size) * (@as(u32, @intFromFloat(w_resized)) / patch_size) / std.math.pow(u32, config.vision_config.spatial_merge_size, 2);
+
+    // Apply the chat template to the prompt
+    const result = try applyChatTemplate(allocator, tokenizer, prompt, number_image_pad_tokens);


result name is reserved for the value you intend to return, so this is misleading.

gwenzek · 2025-12-04T09:27:27Z

examples/qwen3_vl/main.zig

+    // Apply the chat template to the prompt
+    const result = try applyChatTemplate(allocator, tokenizer, prompt, number_image_pad_tokens);
+    const prompt_encoded = result.prompt_tokens;
+    const prompt_shape = result.prompt_shape;


avoid this alias only used once or twice.

gwenzek · 2025-12-04T09:31:14Z

examples/qwen3_vl/main.zig

+    const prompt_shape_buffer = try zml.HostBuffer.empty(allocator, zml.Shape.init(.{ .chw = 3 }, .i32));
+    @memcpy(prompt_shape_buffer.mutItems(i32), &prompt_shape);
+    const image_size_buffer = try zml.HostBuffer.empty(allocator, zml.Shape.init(.{ .chw = 3 }, .i32));
+    @memcpy(image_size_buffer.mutItems(i32), &image_size);


when a variable (like image_size) is only used once, prefer inlining it's definition to where it's used.
In any case keep the definition just above where it's used, so I don't have to scroll up to understand the code.

gwenzek · 2025-12-04T09:35:40Z

examples/qwen3_vl/main.zig

+    @memcpy(prompt_shape_buffer.mutItems(i32), &prompt_shape);
+    const image_size_buffer = try zml.HostBuffer.empty(allocator, zml.Shape.init(.{ .chw = 3 }, .i32));
+    @memcpy(image_size_buffer.mutItems(i32), &image_size);
+    const token_index_buffer = try zml.HostBuffer.empty(allocator, zml.Shape.init(.{}, .i64));


feels like we should have a fromArray that allocates and copy the array content and rename current fromArray to fromArrayPtr

…efacto

This reverts commit 8360600.

Louis Lechevalier and others added 30 commits October 20, 2025 18:50

qwen 3 vl couches de bases pour vision encoder

6a39046

visual + text model

3b9be5c

update for testing

62fb682

update test

cdbf861

push for test

e1286e8

update CLI for path

ca1432e

add test

855dd2e

test

f9240c3

test

4745361

model for test

5a47f14

dump activations bf16

3b18b26

test layers on gpu

911a2c5

update

04f22fa

test vision

518fbfa

test model

cd7f808

update for test

9e37946

test low level layers

c825675

test norm

159a8ac

test on gpu

63dea58

add missing file

4154523

main updated

a7ba72c

stable prefill

7b0696a

qwen with image resize and known seq length

45f3dd0

debug

c48a700

test on gpu

5334652

debug

fdcca7a

remove prints

9eb80b9

remove prints

7614734

remove prints

8208bde

qwen working with a resized image and known seq length and associated…

bcdb249

… main (adding bfactory fork library)

loupicaaa requested review from Corendos, gwenzek and steeve November 25, 2025 15:15

fix for CI

bd2c238

gwenzek reviewed Nov 25, 2025

View reviewed changes

gwenzek requested changes Nov 26, 2025

View reviewed changes

Louis Lechevalier added 4 commits November 26, 2025 18:53

update name

52e8310

push to fix for PR

4d06207

implement test for build position ids

1f49c10

use convolution directly + reusebuffer usage + remove mrope deltas a…

f574571

…s an output

gwenzek requested changes Dec 1, 2025

View reviewed changes

apply modificationsfor pr

feda6c0

gwenzek requested changes Dec 4, 2025

View reviewed changes

Louis Lechevalier added 4 commits December 4, 2025 16:41

updating after PR review

279cdd9

updating existing hostBuffer.fromArray to fromArrayPtr to match the r…

509e93b

…efacto

testing mlp capacity

8360600

Revert "testing mlp capacity"

89c94b9

This reverts commit 8360600.

Qwen 3 vl #359

Are you sure you want to change the base?

Qwen 3 vl #359

Uh oh!

Conversation

loupicaaa commented Nov 25, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gwenzek commented Nov 26, 2025

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gwenzek Nov 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Edit:

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

gwenzek Nov 26, 2025 •

edited

Loading