[Performance] Fuse RoPE + KV cache update for MLA backends by ElizaWszola · Pull Request #35879 · vllm-project/vllm

ElizaWszola · 2026-03-03T16:27:39Z

Follow-up to #32335. Fuse RoPE and MLA KV-cache update into concat_and_cache_mla_rope_fused kernel.

Eval (B200)

lm-eval --model vllm --model_args '{"pretrained": "deepseek-ai/DeepSeek-V2-Lite", "compilation_config": {"use_inductor_graph_partition": "True", "custom_ops": ["+rotary_embedding"]}}' --tasks gsm8k --batch_size auto

with fusion:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.3768	±	0.0133
		strict-match	5	exact_match	↑	0.3738	±	0.0133

no fusion:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.3813	±	0.0134
		strict-match	5	exact_match	↑	0.3783	±	0.0134

Perf (H100, RoPE FlashInfer)

Obtained with

compilation_config=CompilationConfig(
            use_inductor_graph_partition=True,
            custom_ops=["+rotary_embedding"],
)

input-len	prefix-len	output-len	rr	main_ttft	pr_ttft	main_tpot	pr_tpot
256	128	128	1	24.97	24.82	4.90	4.88
512	256	256	1	30.35	30.80	5.71	5.72
1024	512	512	1	88.95	88.55	8.54	8.33
256	128	128	2.5	23.70	23.93	5.99	6.01
512	256	256	2.5	29.02	30.08	8.28	7.91
1024	512	512	2.5	40.18	38.21	11.58	11.13
256	128	128	5	27.62	27.75	7.95	7.59
512	256	256	5	35.11	34.49	10.38	10.07
1024	512	512	5	43.57	41.87	12.79	12.57
256	128	128	10	35.59	35.14	10.43	10.41
512	256	256	10	50.82	46.71	14.30	14.07
1024	512	512	10	57.86	51.05	16.17	15.32
256	128	128	25	43.89	42.95	13.89	13.52
512	256	256	25	85.78	74.02	23.70	21.88
1024	512	512	25	111.77	90.61	40.70	35.09
256	128	128	50	54.36	52.29	19.72	18.91
512	256	256	50	139.89	117.83	47.51	41.59
1024	512	512	50	183.21	156.82	54.22	45.50

Perf (H100, Deepseek Scaling RoPE)

Obtained with

compilation_config=CompilationConfig(
            use_inductor_graph_partition=True,
)

input-len	prefix-len	output-len	rr	main_ttft	pr_ttft	main_tpot	pr_tpot
256	128	128	1	23.71	26.19	4.83	4.91
512	256	256	1	29.06	43.61	5.75	5.77
1024	512	512	1	87.51	85.85	8.33	8.75
256	128	128	2.5	22.86	22.89	5.98	5.83
512	256	256	2.5	27.37	28.61	7.76	7.76
1024	512	512	2.5	37.85	37.74	11.23	10.91
256	128	128	5	26.99	26.44	7.58	7.78
512	256	256	5	33.02	33.76	10.46	10.41
1024	512	512	5	40.59	40.02	12.37	12.28
256	128	128	10	34.86	37.12	10.23	10.72
512	256	256	10	46.38	47.56	14.00	14.04
1024	512	512	10	50.35	50.93	15.22	15.41
256	128	128	25	43.02	45.91	13.84	14.72
512	256	256	25	71.63	79.39	22.51	30.69
1024	512	512	25	98.38	109.80	37.21	48.59
256	128	128	50	53.34	57.97	20.26	26.65
512	256	256	50	118.50	134.55	42.78	45.82
1024	512	512	50	155.14	154.33	46.64	48.71

(The perf results could be improved for Deepseek Scaling RoPE)

Signed-off-by: ElizaWszola <ewszola@redhat.com>

mergify · 2026-03-03T16:28:43Z

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @ElizaWszola.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

gemini-code-assist

Code Review

This pull request refactors the RoPE and KV cache update logic for MLA backends to improve performance with torch.compile. The changes introduce a new custom op unified_mla_kv_cache_update and move the cache update logic into a do_kv_cache_update method on the attention implementation classes. This is a good approach to manage side effects and dependencies for torch.compile. However, I've found a critical issue where a None value for slot_mapping is not handled, which could lead to a runtime crash. Please see my comments for details and suggested fixes.

gemini-code-assist · 2026-03-03T16:30:13Z

+    def do_kv_cache_update(
+        self,
+        kv_c_normed: torch.Tensor,
+        k_pe: torch.Tensor,
+        kv_cache: torch.Tensor,
+        slot_mapping: torch.Tensor,
+        kv_cache_dtype: str,
+        k_scale: torch.Tensor,
+    ) -> None:
+        if kv_cache.numel() == 0:
+            return
+        from vllm import _custom_ops as ops
+
+        ops.concat_and_cache_mla(
+            kv_c_normed,
+            k_pe.squeeze(1),
+            kv_cache,
+            slot_mapping.flatten(),
+            kv_cache_dtype=kv_cache_dtype,
+            scale=k_scale,
+        )


The slot_mapping parameter can be None when slot_mapping.get(self.layer_name) in MLAAttention.forward returns None. This would cause an AttributeError: 'NoneType' object has no attribute 'flatten' when slot_mapping.flatten() is called. You should add a check to handle the case where slot_mapping is None.

Also, the type hint for slot_mapping should be updated to torch.Tensor | None to reflect this.

def do_kv_cache_update( self, kv_c_normed: torch.Tensor, k_pe: torch.Tensor, kv_cache: torch.Tensor, slot_mapping: torch.Tensor | None, kv_cache_dtype: str, k_scale: torch.Tensor, ) -> None: if kv_cache.numel() == 0 or slot_mapping is None: return from vllm import _custom_ops as ops ops.concat_and_cache_mla( kv_c_normed, k_pe.squeeze(1), kv_cache, slot_mapping.flatten(), kv_cache_dtype=kv_cache_dtype, scale=k_scale, )

gemini-code-assist · 2026-03-03T16:30:13Z

+    def do_kv_cache_update(
+        self,
+        kv_c_normed: torch.Tensor,
+        k_pe: torch.Tensor,
+        kv_cache: torch.Tensor,
+        slot_mapping: torch.Tensor,
+        kv_cache_dtype: str,
+        k_scale: torch.Tensor,
+    ) -> None:
+        if kv_cache.numel() == 0:
+            return
+        from vllm import _custom_ops as ops
+
+        ops.concat_and_cache_mla(
+            kv_c_normed,
+            k_pe.squeeze(1),
+            kv_cache,
+            slot_mapping.flatten(),
+            kv_cache_dtype=kv_cache_dtype,
+            scale=k_scale,
+        )


Similar to the do_kv_cache_update in MLAAttentionImpl, slot_mapping can be None. This will cause a crash. Please add a None check for slot_mapping and update its type hint.

def do_kv_cache_update( self, kv_c_normed: torch.Tensor, k_pe: torch.Tensor, kv_cache: torch.Tensor, slot_mapping: torch.Tensor | None, kv_cache_dtype: str, k_scale: torch.Tensor, ) -> None: if kv_cache.numel() == 0 or slot_mapping is None: return from vllm import _custom_ops as ops ops.concat_and_cache_mla( kv_c_normed, k_pe.squeeze(1), kv_cache, slot_mapping.flatten(), kv_cache_dtype=kv_cache_dtype, scale=k_scale, )

Signed-off-by: ElizaWszola <ewszola@redhat.com>

…revisit this Signed-off-by: ElizaWszola <ewszola@redhat.com>

Signed-off-by: ElizaWszola <ewszola@redhat.com>

rbrugaro-amd · 2026-04-22T21:53:58Z

@ElizaWszola , create PR neuralmagic#163 against yours that adds one more fusion to eliminate the rematerialization of k_pe

before (PR35789)

after (this PR)

Accuracy is passing and currently collecting E2E perf on Kimi-k2-FP4
cc: @Rohan138 @attila-dusnoki-htec

Rohan138 · 2026-04-24T18:17:44Z

+    kv_cache_scale: torch.Tensor,
+    layer_name: LayerNameType,
+) -> torch.Tensor:
+    forward_context = get_forward_context()


suggestion:

from vllm.model_executor.layers.attention.attention import get_attention_context ... layer_name = _resolve_layer_name(layer_name) attn_metadata, _, kv_cache, layer_slot_mapping = get_attention_context(layer_name) if layer_slot_mapping is not None: ops.concat_and_cache_mla_rope_fused( positions, q_pe, k_pe, kv_c, cos_sin_cache, is_neox, layer_slot_mapping, kv_cache, kv_cache_dtype, kv_cache_scale, ) return torch.empty(0, device=kv_c.device, dtype=kv_c.dtype)

Slightly more concise, plus you can remove the has_slot_mapping check from the kernel itself

Rohan138 · 2026-04-24T18:21:28Z

+)
+
+
+class KVCacheMLARoPEFusionPattern:


These two classes (KVCacheMLARoPEFusionPattern and KVCacheMLARoPEDeepseekScalingFusionPattern) can be merged with a single use_deepseek_scaling: bool arg, pending additional cleanup in #39488:

if use_deepseek_scaling: self.rope_matcher = MatcherDeepseekScalingRotaryEmbedding( is_neox=self.is_neox, head_size=self.qk_rope_head_dim, num_heads=self.num_heads, num_kv_heads=self.num_kv_heads, ) else: self.rope_matcher = MatcherRotaryEmbedding( # type: ignore is_neox=self.is_neox, head_size=self.qk_rope_head_dim, num_heads=self.num_heads, num_kv_heads=self.num_kv_heads, use_flashinfer=self.use_flashinfer, )

Rohan138 · 2026-04-24T18:24:54Z

        cos, sin = cos_sin.chunk(2, dim=-1)
-        if self.is_neox_style:
+        if is_neox_style:
            # NOTE(woosuk): Here we assume that the positions tensor has the


bugfix: I believe this is a leftover comment from vLLM v0. in V1 this should always be positions.shape == (T,). This should just be:

if is_neox_style: cos = cos.repeat(1, 2).unsqueeze(-2) sin = sin.repeat(1, 2).unsqueeze(-2)

Without this change, adding a unit test with neox_style=True fails with shape errors.

Edit: this was fixed by https://github.com/vllm-project/vllm/pull/40860/changes#diff-eb92572a9c15355e236585bf8c64f01363ab8d04f61385d72ae8696e8568f170

Rohan138 · 2026-04-24T18:30:22Z

  const int64_t block_idx = slot_idx / block_size;
  const int64_t entry_idx = slot_idx % block_size;

  // NOTE: slot_idx can be -1 if the token is padded


We can probably move this early return up before the Q RoPE, if this is a padding token due to V1 cudagraphs

Rohan138 · 2026-04-24T18:37:44Z

+        return [torch.ops.vllm.fused_concat_and_cache_mla_rope.default]
+
+
+@pytest.mark.parametrize(


For ROCm support, can we add:

if current_platform.is_cuda(): MLA_BACKENDS = [ AttentionBackendEnum.FLASH_ATTN_MLA, AttentionBackendEnum.FLASHINFER_MLA, ] elif is_aiter_found_and_supported(): MLA_BACKENDS = [ AttentionBackendEnum.TRITON_MLA, AttentionBackendEnum.ROCM_AITER_MLA, ] else: MLA_BACKENDS = [] @pytest.mark.parametrize("attn_backend", MLA_BACKENDS)

Also, this test should probably be in tests/compile/passes not tests/kernels/core

Rohan138 · 2026-04-24T18:40:08Z

+            "Enabled is only supported with use_deepseek_scaling_rope (and vice versa)"
+        )
+
+    if is_neox and use_deepseek_scaling_rope:


See comments about merging the two patterns and the neox_style fix in deepseek_scaling_rope.py

Rohan138 · 2026-04-24T18:41:38Z

+        fusion_pass = KVCacheMLARoPEFusionPass(vllm_config)
+        passes = [
+            NoOpEliminationPass(vllm_config),
+            SplitCoalescingPass(vllm_config),


don't need SplitCoalescing + ScatterSplitReplacement here, those are only for MHA RoPE model graphs; we don't see those patterns in MLA

Rohan138 · 2026-04-24T18:44:31Z

+        )
+
+
+class KVCacheMLARoPEDeepseekScalingFusionPattern:


suggestion: we could use the new VllmPatternReplacement + VllmPatternMatcherPass here: https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/vllm_inductor_pass.py

Rohan138 · 2026-04-24T18:46:07Z

+def enable_rope_kvcache_mla_fusion(cfg: "VllmConfig") -> bool:
+    """Enable if use_inductor_graph_partition is enabled."""
+
+    return cfg.compilation_config.use_inductor_graph_partition


suggestion, same as enable_rope_kvcache_fusion:

return ( cfg.compilation_config.use_inductor_graph_partition or not cfg.compilation_config.splitting_ops_contain_kv_cache_update() )

Rohan138 · 2026-04-24T18:56:58Z

@ElizaWszola @ProExpertProg I added some comments with the changes from my branch: #40392

Also, one question: Can we use functional torch.aten ops for pattern matching? This PR uses the torch.aten ops for copy, slice_scatter, etc. in the pattern matching, while I was trying to use a pattern like this:

        def _pattern(
            q: torch.Tensor,
            k_pe: torch.Tensor,
            kv_c_normed: torch.Tensor,
            positions: torch.Tensor,
            cos_sin_cache: torch.Tensor,
            k_scale: torch.Tensor,
        ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
            q = q.view(-1, self.num_heads, self.qk_head_dim)
            k_pe_unsqueezed = k_pe.unsqueeze(1)
            q[..., self.qk_nope_head_dim :], k_pe = self.rope_matcher(
                positions,
                q[..., self.qk_nope_head_dim :],
                k_pe_unsqueezed,
                cos_sin_cache,
            )
            dummy = torch.ops.vllm.unified_mla_kv_cache_update(
                kv_c_normed, k_pe, _ln, self.kv_cache_dtype, k_scale
            )
            return dummy, q, k_pe

but this does not work as is and needs additional complexity/preprocessing on the q_slice_scatter op due to the extra slice+view+copy: https://github.com/neuralmagic/vllm/blob/81f2ce414dd41239296f81db0b92200c92456f94/vllm/compilation/passes/fusion/kv_cache_mla_rope_fusion.py#L330

ProExpertProg · 2026-04-25T01:15:41Z

@Rohan138 I'm confused about your question: are you saying we shouldn't be using aten ops? Or are you saying regular torch ops don't work? Or both?

Rohan138 · 2026-04-28T19:47:42Z

@Rohan138 I'm confused about your question: are you saying we shouldn't be using aten ops? Or are you saying regular torch ops don't work? Or both?

Sorry, missed your comment; but yes to both. I realized why I wasn't sure about matching the entire aten sequence in the pattern here-currently, the copy+slice_scatter is introduced during AOTAutograd functionalization, but the replacement completely removes the functionalized write into q. So the graph after the fusion but before defunctionalization doesn't actually write the q_pe back into the base q tensor at all; which I'm not sure if we're fine with?

I got a slightly different torch pattern to work in #40392, with a custom defunctionalization pass:

Pattern:
...
        def _pattern(
            q_pe: torch.Tensor,
            k_pe: torch.Tensor,
            kv_c_normed: torch.Tensor,
            positions: torch.Tensor,
            cos_sin_cache: torch.Tensor,
            k_scale: torch.Tensor,
        ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
            k_pe_unsqueezed = k_pe.unsqueeze(1)
            q_pe, k_pe = self.rope_matcher(
                positions, q_pe, k_pe_unsqueezed, cos_sin_cache
            )
            dummy = torch.ops.vllm.unified_mla_kv_cache_update(
                kv_c_normed, k_pe, _ln, self.kv_cache_dtype, k_scale
            )
            return dummy, q_pe, k_pe

        def _replacement(
            q_pe: torch.Tensor,
            k_pe: torch.Tensor,
            kv_c_normed: torch.Tensor,
            positions: torch.Tensor,
            cos_sin_cache: torch.Tensor,
            k_scale: torch.Tensor,
        ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
            at = auto_functionalized(
                self.FUSED_OP,
                positions=positions,
                q_pe=q_pe,
                k_pe=k_pe,
                kv_c=kv_c_normed,
                cos_sin_cache=cos_sin_cache,
                is_neox=self.is_neox,
                kv_cache_dtype=self.kv_cache_dtype,
                kv_cache_scale=k_scale,
                layer_name=_ln,
            )
            dummy, q_pe, k_pe_squeezed = at
            k_pe = k_pe_squeezed.unsqueeze(1)
            return dummy, q_pe, k_pe
...

Defunctionalization:

...
            elif (
                hasattr(torch.ops.vllm, "fused_rope_unified_mla_kv_cache_update")
                and at_target
                == torch.ops.vllm.fused_rope_unified_mla_kv_cache_update.default
            ):
                # AOTAutograd functionalizes `q[..., nope_dim:] = rope_result` into
                # a sequence of aten ops on q: view+slice+copy+slice_scatter.
                # Since the fused MLA RoPE op mutates q_pe in-place, we can remove
                # the redundant copy and slice_scatter ops during defunctionalization.
                getitem_nodes = self.getitem_users(node)
                q_pe_out = getitem_nodes[1]

                for user in list(q_pe_out.users):
                    if is_func(user, torch.ops.aten.copy.default):
                        copy_temp = user
                slice_temp = user.args[0]
                for user in list(copy_temp.users):
                    if is_func(user, torch.ops.aten.slice_scatter.default):
                        slice_scatter_temp = user
                view_temp = slice_scatter_temp.args[0]

                view_orig = slice_temp.args[0]
                slice_scatter_temp.replace_all_uses_with(view_orig)
                self._remove(slice_scatter_temp)
                self._remove(copy_temp)
                self._remove(slice_temp)
                self._remove(view_temp)
                self._remove(q_pe_out)

                # defunctionalize k_pe manually; self.replace_users_with_mutated_args
                # does not support only replacing specific kwargs
                k_pe_in = node.kwargs["k_pe"]
                k_pe_out = getitem_nodes[2]
                k_pe_out.replace_all_uses_with(k_pe_in)
                self._remove(k_pe_out)

                self.insert_defunctionalized(graph, node)
                self._remove(node)
...

ProExpertProg · 2026-04-29T01:59:07Z

@Rohan138 I think that's actually more correct, I know defunctionalization is complex for rope with the views and slice scatters so it makes sense it would be here as well.

ElizaWszola · 2026-05-12T11:40:34Z

Implemented in #40392

ElizaWszola added 4 commits February 26, 2026 13:29

add dep branch changes

1a2fb76

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Merge branch 'main' into fuse-mla-rope-kv-update

cfa39ec

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Merge branch 'main' into fuse-mla-rope-kv-update

4d92201

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Fusion file

3b7b772

Signed-off-by: ElizaWszola <ewszola@redhat.com>

mergify Bot added the v1 label Mar 3, 2026

mergify Bot added the needs-rebase label Mar 3, 2026

gemini-code-assist Bot reviewed Mar 3, 2026

View reviewed changes

ElizaWszola added 6 commits March 5, 2026 14:36

VIP pattern

63628e5

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Matching working for non-native flashinfer

8fb9530

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Fused runs complete but the output is bad

9c443df

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Placeholder setup to run

0cdf526

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Merge branch 'main' into fuse-mla-rope-kv-update

a1e7366

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Merge branch 'main' into fuse-mla-rope-kv-update

8ff6ca4

Signed-off-by: ElizaWszola <ewszola@redhat.com>

mergify Bot removed the needs-rebase label Mar 11, 2026

ElizaWszola added 15 commits March 11, 2026 15:54

Make kernel not do the cache update when we have no mapping, need to …

bc8d5b1

…revisit this Signed-off-by: ElizaWszola <ewszola@redhat.com>

Working with inductor partitioning, cleanup

4bcf99f

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Cleanup

f8953a5

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Don't specialize kernel for no slot mapping, run static rope instead

8f65aa4

Signed-off-by: ElizaWszola <ewszola@redhat.com>

newline

ad07ad0

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Test fix

049afb8

Signed-off-by: ElizaWszola <ewszola@redhat.com>

More cleanup, formatting

15d281a

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Appease formatter

0dd0dae

Signed-off-by: ElizaWszola <ewszola@redhat.com>

(needs cleanup) get rid of unnecessary shuffles in compiled code

4eb2504

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Cleanup

bcecf4a

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Cleanup enabled codepath

97ad6fa

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Merge branch 'main' into fuse-mla-rope-kv-update

c72291d

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Post-merge fix

4286378

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Avoid extra empty tensor creation in compiled code

12fffd3

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Config variable name, format fix

d7a357c

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Merge branch 'main' into fuse-mla-rope-kv-update

5bf1ea8

Signed-off-by: ElizaWszola <ewszola@redhat.com>

mergify Bot removed the needs-rebase label Apr 13, 2026

ElizaWszola added 2 commits April 15, 2026 05:01

Post-merge fixes (layernametype)

5bb09df

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Get rid of type conversion in fusion

81f2ce4

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Rohan138 mentioned this pull request Apr 20, 2026

[Performance][DSR1]: Fused RoPE+KVCache+q_concat for MLA #40392

Merged

4 tasks

rbrugaro-amd mentioned this pull request Apr 22, 2026

[Perf] Fuse RoPE + KV cache update for MLA backends (#35879): eliminate extra Inductor copy + enable on ROCm neuralmagic/vllm#163

Closed

2 tasks

Rohan138 reviewed Apr 24, 2026

View reviewed changes

vllm-project deleted a comment from aumghelani Apr 25, 2026

xaguilar-amd mentioned this pull request May 3, 2026

[Performance][MLA] Lift decode Q-prep (q-absorb + cat + FP8 quant) out of forward_impl #41568

Open

ElizaWszola closed this May 12, 2026

		return [torch.ops.vllm.fused_concat_and_cache_mla_rope.default]


		@pytest.mark.parametrize(

		)


		class KVCacheMLARoPEFusionPattern:

		)


		class KVCacheMLARoPEDeepseekScalingFusionPattern:

Uh oh!

Conversation

ElizaWszola commented Mar 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Eval (B200)

Perf (H100, RoPE FlashInfer)

Perf (H100, Deepseek Scaling RoPE)

Uh oh!

mergify Bot commented Mar 3, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Mar 3, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Mar 3, 2026

Choose a reason for hiding this comment

Uh oh!

rbrugaro-amd commented Apr 22, 2026

before (PR35789)

after (this PR)

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Rohan138 Apr 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

Rohan138 commented Apr 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ProExpertProg commented Apr 25, 2026

Uh oh!

Rohan138 commented Apr 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ProExpertProg commented Apr 29, 2026

Uh oh!

ElizaWszola commented May 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

ElizaWszola commented Mar 3, 2026 •

edited

Loading

Rohan138 Apr 24, 2026 •

edited

Loading

Rohan138 commented Apr 24, 2026 •

edited

Loading

Rohan138 commented Apr 28, 2026 •

edited

Loading