[Issue 101][eval-and-fix] Guard SeedVR2 var_attention_pytorch on missing nested-jagged API (closes pollockjj/mydevelopment#210)

[pollockjj]

Plan: Guard SeedVR2 var_attention_pytorch against missing torch.nested.nested_tensor_from_jagged

Overview

comfy.ldm.modules.attention.var_attention_pytorch calls torch.nested.nested_tensor_from_jagged unconditionally on lines 731-733 with no hasattr check, no version probe, and no try/except. A user on a PyTorch build that does not provide that API hits a bare AttributeError: module 'torch.nested' has no attribute 'nested_tensor_from_jagged' at the first SeedVR2 inference call — confusing error, no SeedVR2 context. This packet adds a small guard at the top of var_attention_pytorch that detects the missing API and raises a RuntimeError whose message identifies SeedVR2 and the required PyTorch capability, plus a regression module that pins (a) the new error class and message on the missing-API path, (b) the present-API path returning the expected output shape, and (c) malformed-offsets propagating the original torch RuntimeError without SeedVR2 context. The regression module is a structural port of Comfy-Org#119 / Comfy-Org#194's _make_standin + halt-point pattern, with pytest's monkeypatch.delattr(torch.nested, 'nested_tensor_from_jagged') substituting for the post-guard sentinel as the missing-API simulation mechanism.

Diagnosis Summary

This is a PLAN-mode packet. The bug mechanism is fully documented in the issue body's Executive Summary and the Comfy-Org#184 post-mortem (commit 11f3a183, PR Comfy-Org#211). No investigation slice is needed; the fix shape is dictated by the issue body's ## Acceptance Intent cells and the _make_standin precedent from #109 / Comfy-Org#119 / Comfy-Org#194.

§Failure Signature

Three cells, each captured live from the unguarded var_attention_pytorch body on pollockjj/ComfyUI@98df2309 (comfy/ldm/modules/attention.py:722-744). Cell labels A1–A3 are referenced verbatim by AC Pre-fix-fingerprint: sub-bullets.

• Cell A1 — torch.nested.nested_tensor_from_jagged is absent (simulated via monkeypatch.delattr(torch.nested, 'nested_tensor_from_jagged') per GTP-6/GTP-7): var_attention_pytorch(q, k, v, heads, cu_seqlens_q, cu_seqlens_k) raises bare AttributeError("module 'torch.nested' has no attribute 'nested_tensor_from_jagged'") at comfy/ldm/modules/attention.py:731 q = torch.nested.nested_tensor_from_jagged(q, offsets=cu_seqlens_q.long()). The exception class is AttributeError (not RuntimeError) and the message contains neither SeedVR2 nor any SeedVR2-context hint. This is the failure mode the post-mortem identified.
• Cell A2 — torch.nested.nested_tensor_from_jagged is present (the default on torch 2.11.0+cu130, GTP-1): var_attention_pytorch(q, k, v, heads, cu_seqlens_q, cu_seqlens_k) with valid 2-D q, k, v of shape (total_tokens=6, embed_dim=heads*head_dim=16) and matching cu_seqlens_q = cu_seqlens_k = torch.tensor([0, 3, 6], dtype=torch.int32) runs the full computation and returns a tensor of shape (6, 16) (verified live by GTP-3). Pre-guard, this is the canonical SeedVR2-inference shape; the guard MUST NOT false-positive on this path.
• Cell A3 — torch.nested.nested_tensor_from_jagged is present (default), cu_seqlens_q is malformed (off-end against q.shape[0], captured live as the case cu_seqlens_q = torch.tensor([0, 3, 7], dtype=torch.int32) against q.shape[0]==6 per GTP-4): var_attention_pytorch(...) propagates RuntimeError("split_with_sizes expects split_sizes to sum exactly to 6 (input tensor's size at dimension 0), but got split_sizes=[3, 4]") from inside torch.nested.nested_tensor_from_jagged. The RuntimeError originates inside torch (builtins.RuntimeError per GTP-4), the message does NOT contain SeedVR2. Pre-guard this is the legitimate-torch-error path; the guard MUST NOT swallow it as the missing-API case (the SeedVR2-context message MUST NOT replace the original torch message here).

§Failure Boundary

• Triggers Cell A1 — any environment where torch.nested.nested_tensor_from_jagged is unavailable: PyTorch builds that predate the API's introduction; nightly builds where the prototype API is gated behind a build flag; experimental forks that strip the nested-tensor module. The post-mortem identifies this as the "first SeedVR2 inference call" failure mode.
• Triggers Cell A2 — any environment where torch.nested.nested_tensor_from_jagged is present (the slot venv at torch 2.11.0+cu130 per GTP-1, and every supported SeedVR2 production environment).
• Triggers Cell A3 — caller passes malformed cu_seqlens_q or cu_seqlens_k (off-end, non-monotonic, or size-mismatched against q.shape[0]); torch raises the legitimate per-call RuntimeError.
• Does not trigger — sibling attention paths (attention_xformers, attention_sage, attention3_sage, attention_flash, attention_pytorch, attention_split, attention_sub_quad, attention_basic); they do NOT call torch.nested.nested_tensor_from_jagged and are out of scope per ## Non-Goals. The guard is added only inside var_attention_pytorch.

§Root Cause

var_attention_pytorch (defined comfy/ldm/modules/attention.py:722) calls torch.nested.nested_tensor_from_jagged unconditionally three times (lines 731, 732, 733) with no presence check, no try/except, and no SeedVR2-context error message. The hard alias optimized_var_attention = var_attention_pytorch (line 745) does NOT participate in the runtime selector cascade on lines 748-767 (sibling attention paths are selected via if model_management.X_enabled(): optimized_attention = attention_X chains; no equivalent chain exists for var_attention). Consequently a user on a PyTorch build without the API hits a bare AttributeError with no SeedVR2 context. The post-mortem combined has_fallback_path: NO + fallback_pattern_matches_siblings: NO verdicts are the source of this packet.

§Proposed Fix

One surgical change on comfy/ldm/modules/attention.py:

var_attention_pytorch (immediately after def var_attention_pytorch(...) and BEFORE if not skip_reshape:) — add a guard whose effect is: if torch.nested.nested_tensor_from_jagged is absent, raise RuntimeError whose message contains BOTH the substring SeedVR2 and the substring nested_tensor_from_jagged. The guard MUST be placed before the first existing executable line of the function body (currently if not skip_reshape: at line 723) so that no torch operation runs before the API check. Two viable implementation shapes — both pass the AC matchers:

1. hasattr-based entry guard:

   if not hasattr(torch.nested, "nested_tensor_from_jagged"):
       raise RuntimeError(
           "SeedVR2 var_attention_pytorch: torch.nested.nested_tensor_from_jagged "
           "is required by this attention path; the installed PyTorch build "
           "does not provide it"
       )

2. try/except-based call guard wrapping the three torch.nested.nested_tensor_from_jagged calls:

   try:
       q = torch.nested.nested_tensor_from_jagged(q, offsets=cu_seqlens_q.long())
       k = torch.nested.nested_tensor_from_jagged(k, offsets=cu_seqlens_k.long())
       v = torch.nested.nested_tensor_from_jagged(v, offsets=cu_seqlens_k.long())
   except AttributeError as exc:
       raise RuntimeError(
           "SeedVR2 var_attention_pytorch: torch.nested.nested_tensor_from_jagged "
           "is required by this attention path; the installed PyTorch build does "
           "not provide it"
       ) from exc

The exact wording of the guard message is a slicer-implementation detail, but the RuntimeError message MUST contain (a) the substring SeedVR2 and (b) the substring nested_tensor_from_jagged. The slicer MAY choose either implementation shape; the AC matchers are implementation-agnostic because both paths yield the same observable on every cell (Cell A1 → SeedVR2-context RuntimeError; Cell A2 → tensor of shape (6, 16); Cell A3 → torch RuntimeError propagates with split_with_sizes substring and without SeedVR2).

§Verification Strategy

A new test module tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py ports the Comfy-Org#119 / Comfy-Org#194 pattern: an _inputs() helper builds the canonical 2-D (q, k, v) plus matching cu_seqlens_q/cu_seqlens_k per GTP-3; three pytest-style test cells exercise the three Failure-Signature cells; each cell additionally asserts inspect.getsource(var_attention_pytorch) contains the substring raise RuntimeError( whose first occurrence appears BEFORE the first occurrence of nested_tensor_from_jagged (the AST source-pin makes each AC individually diagnostic — pre-fix the unguarded source has no raise RuntimeError( substring, so the source-pin assertion fails for every cell on the base SHA). ACs are evaluated by pytest -q exit code on independent pre-fix and post-fix source checkouts; pre-fix logs MUST exit non-zero (AC-1 cell fails on the matcher mismatch against AttributeError; AC-2 / AC-3 cells fail on the AST source-pin assertion); post-fix logs MUST exit 0 (all three cells pass).

Affected Repositories

Repository	Path	Base Branch	Issue Branch
pollockjj/ComfyUI	/home/johnj/dev_master/ComfyUI	issue_101	issue_210
pollockjj/mydevelopment	/home/johnj/dev_master/mydevelopment	main	issue_210

pollockjj/ComfyUI:issue_210 is created from origin/issue_101 HEAD 98df2309; deliverable PR targets issue_101 via --deliverable-base issue_101. pollockjj/mydevelopment:issue_210 is created from origin/main HEAD 11f3a183; bookkeeping PR targets main and carries the plan, gate outputs, post-mortem, and committed pre/post evidence logs under github_issues/210/slice1/.

Research and Methodology

Plan Foundations Comment URL: https://github.com/pollockjj/mydevelopment/issues/210#issuecomment-4383394124

Detected Scope: none — port of the internal _make_standin + unittest.mock.patch.object halt-point regression pattern established by #109 (tests-unit/comfy_test/seedvr_model_test.py) and reused by Comfy-Org#119 / Comfy-Org#194 (tests-unit/comfy_test/test_diffusers_metadata_guard.py); the existing precedent IS the spec for how this guard's regression module is shaped, and the issue body's Acceptance Intent IS the spec for the production guard text.

The linked comment is the load-bearing artifact for every measurement, equivalence rule, canonical-protocol, tool version, and pipeline-stage citation in this plan. This plan defines no measurable output, no metric, and no equivalence rule against a published anchor; the change class is "API-misuse hygiene guard" with binary RuntimeError-and-substring assertions per AC. No AC below cites a metric, threshold, equivalence rule, or canonical protocol; therefore no AC carries a Foundations-pin: for R&M sub-check 6 (the per-AC scope is N/A across all ACs).

Tools, Pipeline, and Measurements

Plan Foundations Comment URL: https://github.com/pollockjj/mydevelopment/issues/210#issuecomment-4383394124

The linked comment's ## Tools, Pipeline, and Measurements section enumerates: (1) Existing Tooling Inventory with REUSE status for every dep (pytest, comfy.cli_args.args, unittest.mock, pytest.MonkeyPatch, inspect.getsource, torch.nested.nested_tensor_from_jagged, comfy.ldm.modules.attention.var_attention_pytorch); (2) Pipeline = single pytest invocation against tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py (no data pipeline; hygiene-class change); (3) Measurements = N/A (no metric, no anchor, binary assertions only); (4) Single-probe-before-sweep = N/A (3 enumerated cells, not a sweep); (5) Boundary-bracketing-before-verdict = N/A (no characterization verdict; binary observables per cell). No KNOWN-GOOD-REF rows requiring resolution. Every AC below uses pytest as the standard test runner and comfy.ldm.modules.attention as the unmodified-vs-modified surface under test; neither constitutes citing a foundations entry for sub-check-6 purposes per the Phase 3 schema definition (Foundations-pin: is required iff the AC text cites the foundations comment for an equivalence rule, metric definition, canonical protocol, tooling decision, or measurement specification — none of which appear in the AC bodies below). Therefore no AC carries a Foundations-pin: for TPM sub-check 6.

Ground Truth Probes

Probes were run live on 2026-05-05 against pollockjj/ComfyUI@98df2309 (origin/issue_101 HEAD; identical to local 7936e591 for comfy/ldm/modules/attention.py per git diff returning no output). The Python interpreter for every probe is /home/johnj/dev_master/ComfyUI/.venv/bin/python (Python 3.12.13, torch 2.11.0+cu130).

Probe ID	Consumed by AC(s)	Probe command	Observed output (verbatim)
GTP-1	Slice 1 AC-1, AC-2, AC-3	cd /home/johnj/dev_master/ComfyUI && .venv/bin/python -c "import torch; print('torch:', torch.__version__); print('has nested_tensor_from_jagged:', hasattr(torch.nested, 'nested_tensor_from_jagged')); print('callable:', callable(torch.nested.nested_tensor_from_jagged))"	torch: 2.11.0+cu130 / has nested_tensor_from_jagged: True / callable: True
GTP-2	Slice 1 AC-1, AC-2, AC-3	cd /home/johnj/dev_master/ComfyUI && .venv/bin/python -c "import sys; sys.path.insert(0, '.'); from comfy.cli_args import args; import torch; args.cpu = True if not torch.cuda.is_available() else args.cpu; import inspect; import comfy.ldm.modules.attention as A; print('var_attention_pytorch in module:', hasattr(A, 'var_attention_pytorch')); print('signature:', inspect.signature(A.var_attention_pytorch)); print('optimized_var_attention is var_attention_pytorch:', A.optimized_var_attention is A.var_attention_pytorch); print('module path:', A.__file__)"	var_attention_pytorch in module: True / signature: (q, k, v, heads, cu_seqlens_q, cu_seqlens_k, skip_reshape=False, skip_output_reshape=False) / optimized_var_attention is var_attention_pytorch: True / module path: /home/johnj/dev_master/ComfyUI/comfy/ldm/modules/attention.py
GTP-3	Slice 1 AC-2	cd /home/johnj/dev_master/ComfyUI && .venv/bin/python -c "from comfy.cli_args import args; import torch; args.cpu = True if not torch.cuda.is_available() else args.cpu; import comfy.ldm.modules.attention as A; heads, head_dim, total_tokens = 2, 8, 6; embed_dim = heads * head_dim; q = torch.randn(total_tokens, embed_dim); k = torch.randn(total_tokens, embed_dim); v = torch.randn(total_tokens, embed_dim); cu = torch.tensor([0, 3, 6], dtype=torch.int32); out = A.var_attention_pytorch(q, k, v, heads, cu, cu); print('out.shape:', tuple(out.shape)); print('expected: ({}, {})'.format(total_tokens, heads * head_dim))"	out.shape: (6, 16) / expected: (6, 16) (UserWarning about prototype-API printed to stderr; not an exception)
GTP-4	Slice 1 AC-3	cd /home/johnj/dev_master/ComfyUI && .venv/bin/python -c "from comfy.cli_args import args; import torch; args.cpu = True if not torch.cuda.is_available() else args.cpu; import comfy.ldm.modules.attention as A; q = torch.randn(6, 16); k = torch.randn(6, 16); v = torch.randn(6, 16); cu_q = torch.tensor([0, 3, 7], dtype=torch.int32); cu_k = torch.tensor([0, 3, 6], dtype=torch.int32); \ntry:\n A.var_attention_pytorch(q, k, v, 2, cu_q, cu_k)\n print('NO RAISE')\nexcept Exception as e:\n print('Exception type:', type(e).__name__); print('Exception module:', type(e).__module__); print('Exception msg head:', str(e)[:200])"	Exception type: RuntimeError / Exception module: builtins / Exception msg head: split_with_sizes expects split_sizes to sum exactly to 6 (input tensor's size at dimension 0), but got split_sizes=[3, 4]
GTP-5	Slice 1 AC-1	cd /home/johnj/dev_master/ComfyUI && .venv/bin/python -c "import torch.nested; print('before:', hasattr(torch.nested, 'nested_tensor_from_jagged')); saved = torch.nested.nested_tensor_from_jagged; delattr(torch.nested, 'nested_tensor_from_jagged'); print('after delattr:', hasattr(torch.nested, 'nested_tensor_from_jagged')); torch.nested.nested_tensor_from_jagged = saved; print('after restore:', hasattr(torch.nested, 'nested_tensor_from_jagged'))"	before: True / after delattr: False / after restore: True
GTP-6	Slice 1 AC-1	cd /home/johnj/dev_master/ComfyUI && .venv/bin/python -c "import _pytest.monkeypatch as mp; import torch.nested; m = mp.MonkeyPatch(); print('before:', hasattr(torch.nested, 'nested_tensor_from_jagged')); m.delattr(torch.nested, 'nested_tensor_from_jagged'); print('after delattr:', hasattr(torch.nested, 'nested_tensor_from_jagged')); m.undo(); print('after undo:', hasattr(torch.nested, 'nested_tensor_from_jagged'))"	before: True / after delattr: False / after undo: True
GTP-7	Slice 1 AC-1, AC-2, AC-3	cd /home/johnj/dev_master/ComfyUI && .venv/bin/python -c "from comfy.cli_args import args; import torch; args.cpu = True if not torch.cuda.is_available() else args.cpu; import inspect; import comfy.ldm.modules.attention as A; src = inspect.getsource(A.var_attention_pytorch); print('full source length:', len(src), 'chars'); print('first nested_tensor_from_jagged at index:', src.index('nested_tensor_from_jagged')); print('contains raise RuntimeError(:', 'raise RuntimeError(' in src)"	full source length: 996 chars / first nested_tensor_from_jagged at index: 445 / contains raise RuntimeError(: False
GTP-8	Slice 1 (informational; deliverable-base SHA pin)	cd /home/johnj/dev_master/ComfyUI && git rev-parse origin/issue_101	98df2309dc91c8de346fa74579d50c54595fc6d8
GTP-9	Slice 1 (informational; mydevelopment-base SHA pin)	cd /home/johnj/dev_master/mydevelopment && git rev-parse origin/main	11f3a18391c2d048b2a747a6aa9d837129f2f87e
GTP-10	Slice 1 AC-2, AC-3 (downstream call-site context — informational)	cd /home/johnj/dev_master/ComfyUI && git show origin/issue_101:comfy/ldm/seedvr/model.py | grep -n -E "var_attention|optimized_var_attention|nested_tensor_from_jagged" | head -10	11:from comfy.ldm.modules.attention import optimized_var_attention / 807: out = optimized_var_attention(

Created Surface Contract

The plan introduces literals that do not exist on the base branch. These are anchored here (not in ## Ground Truth Probes) because they will be created by Slice 1 itself; their post-creation existence is verified by the slice's own ACs.

Created literal	Created by	Verified by
tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py (new test module)	Slice 1 implementation commit on pollockjj/ComfyUI:issue_210	AC-1, AC-2, AC-3 all assert via this module's pytest invocations
_inputs (helper in new test module producing the canonical (q, k, v, heads, cu_seqlens_q, cu_seqlens_k, total_tokens, embed_dim) tuple per GTP-3)	Slice 1	AC-1, AC-2, AC-3 (consumed by all behavioral cells)
test_missing_api_raises_seedvr2_runtime_error (test fn — AC-1 cell; uses monkeypatch.delattr(torch.nested, 'nested_tensor_from_jagged') per GTP-6 to simulate the absent-API state)	Slice 1	AC-1
test_present_api_returns_expected_shape (test fn — AC-2 cell; calls var_attention_pytorch with valid inputs and asserts out.shape == (total_tokens, heads * head_dim))	Slice 1	AC-2
test_malformed_offsets_propagates_torch_runtime_error (test fn — AC-3 cell; calls var_attention_pytorch with off-end cu_seqlens_q = torch.tensor([0, 3, 7], dtype=torch.int32) per GTP-4, asserts torch RuntimeError propagates with split_with_sizes substring AND SeedVR2 substring is NOT in the message)	Slice 1	AC-3
AST source-pin assertion (in each of AC-1/AC-2/AC-3 cells): inspect.getsource(var_attention_pytorch) contains raise RuntimeError( whose first occurrence appears BEFORE the first occurrence of nested_tensor_from_jagged. This pin is the per-AC diagnostic anchor: pre-fix the source contains nested_tensor_from_jagged but no raise RuntimeError( (GTP-7 confirms contains raise RuntimeError(: False); the assertion fails on the base SHA. Post-fix the new guard's raise RuntimeError( substring appears before the first existing nested_tensor_from_jagged call.	Slice 1	AC-1, AC-2, AC-3
Production guard text in comfy/ldm/modules/attention.py var_attention_pytorch (1 guard, raising RuntimeError whose message contains substring SeedVR2 AND nested_tensor_from_jagged; placed before the function's existing first executable line if not skip_reshape: at base-line 723)	Slice 1 implementation commit on pollockjj/ComfyUI:issue_210	AC-1 (behavioral) + AC-1/AC-2/AC-3 (AST source-pin component)
Pre-fix log artifacts at github_issues/210/slice1/<test>_pre.log (one per AC)	Slice 1 (slicer runs pytest against worktree at 98df2309 with the new test file dropped in but the production guard absent)	AC-1, AC-2, AC-3 cite the file paths verbatim
Post-fix log artifacts at github_issues/210/slice1/<test>_post.log (one per AC)	Slice 1 (slicer runs pytest against pollockjj/ComfyUI:issue_210 with both the production guard and the test module applied)	AC-1, AC-2, AC-3 cite the file paths verbatim

Asset Readiness Inventory

Asset	Location	Provenance	Status
comfy/ldm/modules/attention.py (unguarded base)	pollockjj/ComfyUI@98df2309:comfy/ldm/modules/attention.py	origin/issue_101 HEAD; PR #46 merge	VERIFIED — exists, opens cleanly, var_attention_pytorch body at lines 722-744 matches §Failure Signature cell sites at 731-733 (probes GTP-2 / GTP-7); pre-fix opaque-AttributeError reproduced live in Phase 0.5 dry-run via GTP-5/GTP-6
tests-unit/comfy_test/seedvr_model_test.py (#109 precedent)	pollockjj/ComfyUI@98df2309:tests-unit/comfy_test/seedvr_model_test.py	merged via PR #33	VERIFIED — file exists on issue_101; _make_standin borrow pattern at function _make_standin(positive_conditioning) is the structural template for this packet's helper layout
tests-unit/comfy_test/test_diffusers_metadata_guard.py (Comfy-Org#119 / Comfy-Org#194 precedent — primary structural template)	pollockjj/ComfyUI@98df2309:tests-unit/comfy_test/test_diffusers_metadata_guard.py	merged via PR #36	VERIFIED — file exists on issue_101; _make_standin + unittest.mock.patch.object halt-point pattern is the structural template (this packet substitutes monkeypatch.delattr for the post-guard sentinel as the missing-API simulation mechanism)
Python 3.12 venv at /home/johnj/dev_master/ComfyUI/.venv	Linux slot dev_master on prosoche	Maintained by Boss; same venv every prior issue_101 packet ran against	VERIFIED — torch 2.11.0+cu130 with torch.nested.nested_tensor_from_jagged callable per GTP-1; pytest.MonkeyPatch.delattr/undo cycle works per GTP-6
tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py (new test module)	pollockjj/ComfyUI:issue_210:tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py	created by Slice 1	BROKEN-EXPECTED — does not exist on base SHA (git ls-tree origin/issue_101 -- tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py returns no output); created by Slice 1 implementation commit; AC-1/AC-2/AC-3 verify post-creation behavior
Production guard in var_attention_pytorch	pollockjj/ComfyUI:issue_210:comfy/ldm/modules/attention.py	created by Slice 1	BROKEN-EXPECTED — guard absent on base SHA (the entire defect being fixed); GTP-7 confirms contains raise RuntimeError(: False; created by Slice 1 implementation commit; AC-1/AC-2/AC-3 verify post-fix behavior via behavioral matchers and AST source-pin
Pre-fix and post-fix evidence logs	pollockjj/mydevelopment:issue_210:github_issues/210/slice1/<test>_{pre,post}.log	created by Slice 1	BROKEN-EXPECTED — written by slicer during Slice 1 evidence capture; AC-1/AC-2/AC-3 cite the paths and the slicer commits the logs

Slices

---

Slice 1: Convert opaque AttributeError on missing torch.nested.nested_tensor_from_jagged into a SeedVR2-named RuntimeError

Kind: implementation

Objective: Prove that var_attention_pytorch raises a RuntimeError whose message contains both SeedVR2 and nested_tensor_from_jagged when torch.nested.nested_tensor_from_jagged is absent (the missing-API path), AND that the present-API path returns the expected output shape (the canonical SeedVR2-inference path is unchanged), AND that malformed cu_seqlens propagates the original torch RuntimeError without SeedVR2-context substitution (the legitimate-torch-error path is unchanged).

Acceptance Criteria

• AC-1: cd /home/johnj/dev_master/ComfyUI && python3 -m pytest -q tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py::test_missing_api_raises_seedvr2_runtime_error exits 0 on the post-fix pollockjj/ComfyUI:issue_210 checkout AND exits non-zero on the pre-fix origin/issue_101 (98df2309) worktree with the same test module dropped in (the test's pytest.raises(RuntimeError, match=r"SeedVR2.*nested_tensor_from_jagged") matcher fails to match the bare AttributeError from the unguarded torch.nested.nested_tensor_from_jagged lookup; AND the AST source-pin assertion 'raise RuntimeError(' in inspect.getsource(var_attention_pytorch) fails because the unguarded source contains no such substring per GTP-7) — verified by committed artifacts github_issues/210/slice1/test_missing_api_raises_seedvr2_runtime_error_pre.log AND github_issues/210/slice1/test_missing_api_raises_seedvr2_runtime_error_post.log, each containing the verbatim pytest invocation, full stdout, and final line exit code: <N>.

  • Pre-fix-fingerprint: on origin/issue_101@98df2309 the unguarded var_attention_pytorch body has no presence check; with monkeypatch.delattr(torch.nested, 'nested_tensor_from_jagged') removing the attribute (per GTP-5 / GTP-6 confirming delattr succeeds and hasattr returns False), the call site at comfy/ldm/modules/attention.py:731 raises bare AttributeError("module 'torch.nested' has no attribute 'nested_tensor_from_jagged'"); the matcher requires RuntimeError and does NOT accept AttributeError (the two classes are unrelated in the exception hierarchy); the test FAILS with the AttributeError propagating through pytest.raises(RuntimeError). Additionally, the AST source-pin assertion fails because per GTP-7 the unguarded source has contains raise RuntimeError(: False (Diagnosis Summary §Failure Signature Cell A1).
  • Post-fix-expectation: post-fix var_attention_pytorch raises RuntimeError whose str(exc.value) contains both substrings SeedVR2 AND nested_tensor_from_jagged (regardless of whether the slicer chose the hasattr-based entry guard or the try/except-based call guard per §Proposed Fix); the matcher matches; AND inspect.getsource(var_attention_pytorch) contains raise RuntimeError( whose first occurrence appears BEFORE the first occurrence of nested_tensor_from_jagged; the test PASSES with exit 0.
  • Assumes-passed: none
  • In-scope-of: Slice 1 Objective "Prove that var_attention_pytorch raises a RuntimeError whose message contains both SeedVR2 and nested_tensor_from_jagged when torch.nested.nested_tensor_from_jagged is absent (the missing-API path), AND that the present-API path returns the expected output shape (the canonical SeedVR2-inference path is unchanged), AND that malformed cu_seqlens propagates the original torch RuntimeError without SeedVR2-context substitution (the legitimate-torch-error path is unchanged)."
  • Probe: GTP-1, GTP-2, GTP-5, GTP-6, GTP-7, GTP-8

• AC-2: cd /home/johnj/dev_master/ComfyUI && python3 -m pytest -q tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py::test_present_api_returns_expected_shape exits 0 on the post-fix pollockjj/ComfyUI:issue_210 checkout AND exits non-zero on the pre-fix origin/issue_101 (98df2309) worktree with the same test module dropped in — the test asserts BOTH (a) var_attention_pytorch(q, k, v, heads=2, cu_seqlens_q=torch.tensor([0,3,6], dtype=torch.int32), cu_seqlens_k=torch.tensor([0,3,6], dtype=torch.int32)) with valid 2-D q, k, v of shape (6, 16) returns a tensor of out.shape == (6, 16) (matching (total_tokens, heads * head_dim) per §Proposed Fix Cell A2 + GTP-3); AND (b) the AST source-pin assertion 'raise RuntimeError(' in inspect.getsource(var_attention_pytorch) and inspect.getsource(var_attention_pytorch).index('raise RuntimeError(') < inspect.getsource(var_attention_pytorch).index('nested_tensor_from_jagged') holds. On the pre-fix base SHA assertion (a) passes (the unguarded computation already produces shape (6, 16) per GTP-3) but assertion (b) FAILS because the unguarded source has no raise RuntimeError( substring per GTP-7; the conjunction fails → exit non-zero — verified by committed artifacts github_issues/210/slice1/test_present_api_returns_expected_shape_pre.log AND github_issues/210/slice1/test_present_api_returns_expected_shape_post.log.

  • Pre-fix-fingerprint: on origin/issue_101@98df2309 inspect.getsource(var_attention_pytorch) does NOT contain the substring raise RuntimeError( (probe GTP-7 captured contains raise RuntimeError(: False on the live module); the AST source-pin conjoined assertion FAILS with AssertionError. The behavioral shape assertion (a) passes vacuously (the unguarded computation already produces the correct shape per GTP-3), but the conjoint test FAILS at the source-pin step (Diagnosis Summary §Failure Signature Cell A2 + §Root Cause).
  • Post-fix-expectation: post-fix inspect.getsource(var_attention_pytorch) contains raise RuntimeError( whose first occurrence appears BEFORE the first occurrence of nested_tensor_from_jagged; the AST source-pin assertion PASSES; AND the behavioral shape assertion (a) continues to hold (the new guard does NOT false-positive on the present-API path because torch.nested.nested_tensor_from_jagged is callable on the slot venv per GTP-1, so the guard's hasattr check or try/except returns control to the original computation); the conjoint test PASSES with exit 0.
  • Assumes-passed: none
  • In-scope-of: Slice 1 Objective "Prove that var_attention_pytorch raises a RuntimeError whose message contains both SeedVR2 and nested_tensor_from_jagged when torch.nested.nested_tensor_from_jagged is absent (the missing-API path), AND that the present-API path returns the expected output shape (the canonical SeedVR2-inference path is unchanged), AND that malformed cu_seqlens propagates the original torch RuntimeError without SeedVR2-context substitution (the legitimate-torch-error path is unchanged)."
  • Probe: GTP-1, GTP-2, GTP-3, GTP-7, GTP-8

• AC-3: cd /home/johnj/dev_master/ComfyUI && python3 -m pytest -q tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py::test_malformed_offsets_propagates_torch_runtime_error exits 0 on the post-fix pollockjj/ComfyUI:issue_210 checkout AND exits non-zero on the pre-fix origin/issue_101 (98df2309) worktree with the same test module dropped in — the test asserts BOTH (a) calling var_attention_pytorch(q, k, v, heads=2, cu_seqlens_q=torch.tensor([0,3,7], dtype=torch.int32), cu_seqlens_k=torch.tensor([0,3,6], dtype=torch.int32)) (off-end cu_seqlens_q[-1]==7 against q.shape[0]==6 per GTP-4) raises RuntimeError whose str(exc.value) contains the substring split_with_sizes AND 'SeedVR2' not in str(exc.value) (the original torch error propagates without SeedVR2-context substitution); AND (b) the AST source-pin assertion 'raise RuntimeError(' in inspect.getsource(var_attention_pytorch) and inspect.getsource(var_attention_pytorch).index('raise RuntimeError(') < inspect.getsource(var_attention_pytorch).index('nested_tensor_from_jagged') holds. On the pre-fix base SHA assertion (a) passes (the unguarded computation already propagates the torch RuntimeError per GTP-4 with no SeedVR2 substring anywhere in the codebase), but assertion (b) FAILS because the unguarded source has no raise RuntimeError( substring per GTP-7; the conjunction fails → exit non-zero — verified by committed artifacts github_issues/210/slice1/test_malformed_offsets_propagates_torch_runtime_error_pre.log AND github_issues/210/slice1/test_malformed_offsets_propagates_torch_runtime_error_post.log.

  • Pre-fix-fingerprint: on origin/issue_101@98df2309 inspect.getsource(var_attention_pytorch) does NOT contain the substring raise RuntimeError( (probe GTP-7); the AST source-pin conjoined assertion FAILS with AssertionError. The behavioral assertion (a) passes vacuously (the unguarded computation already propagates RuntimeError("split_with_sizes expects split_sizes to sum exactly to 6...") from torch per GTP-4, and the substring SeedVR2 is not in any pre-fix code path), but the conjoint test FAILS at the source-pin step (Diagnosis Summary §Failure Signature Cell A3 + §Root Cause).
  • Post-fix-expectation: post-fix inspect.getsource(var_attention_pytorch) contains raise RuntimeError( whose first occurrence appears BEFORE the first occurrence of nested_tensor_from_jagged; the AST source-pin assertion PASSES; AND the behavioral assertion (a) continues to hold — calling var_attention_pytorch with off-end cu_seqlens_q propagates the torch RuntimeError containing split_with_sizes substring, AND 'SeedVR2' not in str(exc.value) because the new guard fires only on the missing-API path (AttributeError from torch.nested.nested_tensor_from_jagged lookup), not on torch's own per-call shape errors; the conjoint test PASSES with exit 0.
  • Assumes-passed: none
  • In-scope-of: Slice 1 Objective "Prove that var_attention_pytorch raises a RuntimeError whose message contains both SeedVR2 and nested_tensor_from_jagged when torch.nested.nested_tensor_from_jagged is absent (the missing-API path), AND that the present-API path returns the expected output shape (the canonical SeedVR2-inference path is unchanged), AND that malformed cu_seqlens propagates the original torch RuntimeError without SeedVR2-context substitution (the legitimate-torch-error path is unchanged)."
  • Probe: GTP-1, GTP-2, GTP-4, GTP-7, GTP-8

Constraints

• Python: python3 (resolved from cd /home/johnj/dev_master/ComfyUI to the slot venv at .venv/bin/python, Python 3.12.13)
• Runner: python3 -m pytest -q tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py (no debug harness; no debug/run_debug.py invocation; this is a unit-test-only packet)
• Isolation flags: N/A (no --use-process-isolation, no --disable-cuda-malloc; the test imports var_attention_pytorch directly from comfy.ldm.modules.attention with comfy.cli_args.args.cpu = True set before import per Add SeedVR2 support #109/UserWarning bug ? Comfy-Org/ComfyUI#119/Use a global random seed for all samplers Comfy-Org/ComfyUI#194 precedent)
• No packages installed into host venv (every dep is REUSE — pytest, stdlib, torch, comfy module already present on the slot venv at /home/johnj/dev_master/ComfyUI/.venv)
• No pkill, no rm -rf, no python main.py
• All commits on issue branch issue_210, not on base branches: deliverable commit on pollockjj/ComfyUI:issue_210 (created from origin/issue_101@98df2309); bookkeeping commit on pollockjj/mydevelopment:issue_210 (created from origin/main@11f3a183)
• Relevant Existing Tooling:
  • pytest (REUSE) — same harness as Add SeedVR2 support #109/UserWarning bug ? Comfy-Org/ComfyUI#119/Use a global random seed for all samplers Comfy-Org/ComfyUI#194 precedent regression modules.
  • comfy.cli_args.args (REUSE) — args.cpu = True toggle for CPU-only test runs; same shape as Add SeedVR2 support #109/UserWarning bug ? Comfy-Org/ComfyUI#119/Use a global random seed for all samplers Comfy-Org/ComfyUI#194 precedent.
  • unittest.mock + pytest.MonkeyPatch (REUSE) — stdlib + pytest-builtin; monkeypatch.delattr(torch.nested, 'nested_tensor_from_jagged') substitutes for UserWarning bug ? Comfy-Org/ComfyUI#119/Use a global random seed for all samplers Comfy-Org/ComfyUI#194's unittest.mock.patch.object(...) halt-point sentinel as the missing-API simulation mechanism (per GTP-6).
  • inspect.getsource (REUSE) — stdlib; AST source-pin assertions per Add SeedVR2 support #109 precedent's inspect.getsource(NaDiT.forward) shape.
  • torch.nested.nested_tensor_from_jagged (REUSE) — the existing API the unguarded code path calls; GTP-1 confirms callable on the slot venv at torch 2.11.0+cu130.
  • comfy.ldm.modules.attention.var_attention_pytorch (REUSE) — the production function under modification; GTP-2 confirms signature and aliasing to optimized_var_attention.
• Evidence Primitive Requirements:
  • Behavioral-change claim (AC-1 missing-API path): independent pre-fix and post-fix source checkouts plus committed pytest log artifacts — AC-1's _pre.log AND _post.log are required.
  • Regression-coverage claims (AC-2, AC-3): same pre-fix / post-fix log shape with the AST source-pin assertion as the diagnostic anchor — _pre.log shows source-pin failure at base SHA, _post.log shows source-pin pass after the guard is added.
• Repository Hygiene:
  • Every dirty repo state is resolved by the slicer without git stash or external decision deferral.
  • Dirty paths are classified as committed work, precise committed .gitignore rules for safe generated bulk, or discarded generated debris.
  • Clean committed work already present on the active branch is repo provenance and is integrated, not removed as dirty-state cleanup.
  • Slicers and Melian must run live git status --short gates locally and must refuse to submit while any touched repo is dirty.
  • Plans must not require repo_hygiene.txt, raw git status --short transcripts, or CLEAN markers as QA evidence.
  • QA verifies submitted commits, branch refs, and artifact presence through GitHub rather than trusting slicer-authored cleanliness transcripts.

Out of Scope

• Any change to comfy/ldm/seedvr/model.py or the call site at :807 (the import of optimized_var_attention and its invocation are unchanged; only var_attention_pytorch body in comfy/ldm/modules/attention.py is modified).
• Any change to sibling attention paths (attention_xformers, attention_sage, attention3_sage, attention_flash, attention_pytorch, attention_split, attention_sub_quad, attention_basic).
• Any rewrite of var_attention_pytorch's math (the nested_tensor_from_jagged calls, the transpose / comfy.ops.scaled_dot_product_attention / reshape sequence, and the skip_reshape / skip_output_reshape parameter handling are preserved verbatim).
• Any new variable-length-attention implementation, selector cascade, or fallback math path. None exist in comfy/ldm/modules/attention.py to select toward; adding one is out of scope per the issue body's ## Non-Goals.
• Editing the #184 review document's verdicts or the post-mortem in PR Is it possible to cancel caching of the final result to to generate with the same parameters? Comfy-Org/ComfyUI#211. The historical record stays as-is; the new behaviour is recorded in this issue's own evidence files under github_issues/210/slice1/.
• Any change to optimized_var_attention (the alias at attention.py:745 remains a hard alias to var_attention_pytorch; no selector cascade is added).
• Any benchmarking, calibration sweep, or measurement-instrument validation. The change class is hygiene-only; no metric, no anchor, no published-baseline comparison.
• PR-review-cycle accommodations beyond the standing pr-review cycle-Sync with Upstream #1+ rule (pre-existing-code Copilot findings such as docstring grammar elsewhere in attention.py are ignorable; only newly-changed lines warrant disposition).

---

Tracked in pollockjj/mydevelopment Comfy-Org#210 (bookkeeping PR pollockjj/mydevelopment#213).

[qa-agent-seveneves]

Codex Review — Cycle 1

Overall: patch is correct
Overall confidence: 0.86
Explanation: No blocking issues were found in the changed lines. The guard is scoped to the missing nested API path, preserves the existing present-API and malformed-offset behavior, and the targeted regression module passes under the repository venv.

No findings.

[Copilot]

Pull request overview

This PR adds a SeedVR2-specific runtime guard to comfy.ldm.modules.attention.var_attention_pytorch so environments lacking torch.nested.nested_tensor_from_jagged fail with a clearer, SeedVR2-context RuntimeError, and it introduces unit tests to lock in the new behavior (missing-API error path, present-API output shape, and malformed-offset error propagation).

Changes:

• Add an early guard in var_attention_pytorch that raises a SeedVR2-named RuntimeError when nested_tensor_from_jagged is not available.
• Add a new regression test module covering missing API, normal execution shape, and malformed-offset propagation, plus a source-introspection pin.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
comfy/ldm/modules/attention.py	Adds a top-of-function capability check and SeedVR2-context error message for missing torch.nested.nested_tensor_from_jagged.
tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py	Adds pytest regressions for the guard behavior and a source-level pin to ensure the guard stays ahead of the nested-tensor calls.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[pollockjj]

INVALID — the production guard already handles torch.nested itself being missing. comfy/ldm/modules/attention.py:730 uses _nested = getattr(torch, 'nested', None) and if _nested is None or not hasattr(_nested, _VAR_ATTENTION_NESTED_API_NAME):, so the torch.nested access cannot raise AttributeError before the check. Test test_missing_namespace_raises_seedvr2_runtime_error (line 114) pins this exact case and passes.

[pollockjj]

INVALID — both delattr calls in this test (lines 105 and 115) pass raising=False, which silently no-ops if the attribute is already absent. That is precisely the safety behaviour Copilot is asking for; pytest's monkeypatch documents raising=False as the suppression flag for missing-attribute teardown.

[pollockjj]

INVALID — _assert_guard_source_pin() is AST-based, not substring-based. Lines 74-101 walk ast.parse(textwrap.dedent(inspect.getsource(var_attention_pytorch))) for ast.Raise (with ast.Call of ast.Name == 'RuntimeError') and ast.Attribute (with .attr == 'nested_tensor_from_jagged'), then asserts min(raise_lines) < min(nested_lines). Refactors that split the raise across lines or move the literal into a docstring or comment do not affect the AST walk. The function docstring (lines 60-73) explicitly explains why the AST approach was chosen over substring positions.

[pollockjj]

NOT-REQUIRED — issue Comfy-Org#210's Acceptance criterion 1 is an explicit OR: "a backend selector is wired ... OR an explicit try: nested_tensor_from_jagged(...) / except AttributeError, RuntimeError guard is added". This PR delivers the guard branch of that OR. Updating the PR title to reflect actual delivery.

[tdd-agent]

Melian PR-Review — Cycle 1 — DECISION: CONTINUE — blocking findings remain.

HEAD SHA reviewed: 9a508457ee9f
Verdict: HOLD

Why we are continuing

4 blocking finding(s) remain. The slicer will address them and the next cycle will re-review the resulting head SHA.

Blocking findings

• [copilot] [P?] The guard uses hasattr(torch.nested, ...), but if a PyTorch build/fork does not expose torch.nested at all, accessin — comfy/ldm/modules/attention.py
  • Block reason: unknown outcome: ''
  • Comment: [Issue 101][eval-and-fix] Guard SeedVR2 var_attention_pytorch on missing nested-jagged API (closes pollockjj/mydevelopment#210) #48 (comment)
• [copilot] [P?] monkeypatch.delattr(torch.nested, 'nested_tensor_from_jagged') will itself raise if the attribute is already missing ( — tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py
  • Block reason: unknown outcome: ''
  • Comment: [Issue 101][eval-and-fix] Guard SeedVR2 var_attention_pytorch on missing nested-jagged API (closes pollockjj/mydevelopment#210) #48 (comment)
• [copilot] [P?] _assert_guard_source_pin() is fairly brittle: it relies on substring positions for raise RuntimeError( and nested_t — tests-unit/comfy_test/test_var_attention_pytorch_seedvr2_guard.py`
  • Block reason: unknown outcome: ''
  • Comment: [Issue 101][eval-and-fix] Guard SeedVR2 var_attention_pytorch on missing nested-jagged API (closes pollockjj/mydevelopment#210) #48 (comment)
• [copilot] [P?] The PR title mentions “Add backend selector + fallback to optimized_var_attention”, but the changes here only add a Seed — comfy/ldm/modules/attention.py
  • Block reason: unknown outcome: ''
  • Comment: [Issue 101][eval-and-fix] Guard SeedVR2 var_attention_pytorch on missing nested-jagged API (closes pollockjj/mydevelopment#210) #48 (comment)

Raw signal

• codex: overall_correctness='patch is correct', confidence=0.86, findings=0
• copilot: state='COMMENTED', inline_comments=4

Verdict rationale (raw)

both reviewers satisfied

This comment is posted by Melian's FSM (_handle_pr_review_triage) on every triage that results in a STOP or CONTINUE decision so the audit trail is complete.

[qa-agent-seveneves]

Codex Review -- Round 1 -- KICKOFF

Runner: scripts/run_codex_review.py (structured-output mode)
PR: #48
Base branch under review: issue_101
Reviewer: codex exec --output-schema codex_review_schema.json (JSON schema enforced server-side)
Kickoff timestamp (UTC): 2026-05-05T23:39:36Z

Awaiting codex output. The structured result will be posted as a separate comment on this PR when codex exits.

[qa-agent-seveneves]

Codex Review -- Round 1 -- RESULT

PR: #48
Base branch: issue_101
Reviewer: codex exec --output-schema (structured output)
Result timestamp (UTC): 2026-05-05T23:40:37Z
Duration (s): 58.9
Verdict: ✅ patch is correct (zero P0/P1/P2 findings)
Confidence: 0.89

Overall explanation

The PR adds the missing SeedVR2 guard before any nested-jagged calls and preserves the present-API and malformed-offset behavior. The targeted regression module passes at PR head with 4 tests passing, so I found no blocking correctness issues.

Findings (0)

No findings.

---

Raw JSON (schema-enforced)

{
  "findings": [],
  "overall_confidence_score": 0.89,
  "overall_correctness": "patch is correct",
  "overall_explanation": "The PR adds the missing SeedVR2 guard before any nested-jagged calls and preserves the present-API and malformed-offset behavior. The targeted regression module passes at PR head with 4 tests passing, so I found no blocking correctness issues."
}