[Issue 101][eval-and-fix] Audit forward-path device and dtype casts (closes pollockjj/mydevelopment#185)#49
Merged
Conversation
…closes pollockjj/mydevelopment#185) Kosinkadink's PR Comfy-Org#11294 review (comment 4259303182) flagged five forward-path device-override sites in comfy/ldm/seedvr/model.py. Yousef removed four; MMModule.forward retained the residual: device = comfy.model_management.get_torch_device() vid = vid.to(device) This pattern fights Comfy's model-management system: load_device already places parameters on a chosen device, and inputs should arrive on the same device. The global get_torch_device() call overrides whatever placement Comfy decided. The cast is removed; comfy.model_management is no longer referenced in this file (import dropped). The surviving line below the removed cast (txt = txt.to(device=vid.device, dtype=vid.dtype)) follows vid.device — Comfy-native pattern, identical to the one NaMMRotaryEmbedding3d.forward already uses. Regression coverage in tests-unit/comfy_test/test_seedvr_forward_no_device_cast.py pins both layers of the fix: 1. Source-level (AST-walked): no forward method body in comfy/ldm/seedvr/model.py calls get_torch_device(); the file no longer references comfy.model_management at all. 2. Smoke forward: MMModule(nn.Linear, ...) runs on a CUDA device and outputs stay on the input device — under shared_weights, vid_only, and standard branches. 74 pre-existing SeedVR2 tests remain bit-identical pass.
…lign tests to contract
Restore the top-level ``import comfy.model_management`` so the diff
between origin/issue_101 and issue_185 contains exactly the two cast
lines inside ``MMModule.forward`` (lines 624-625 on origin/issue_101)
and nothing else. Rewrite ``test_seedvr_forward_no_device_cast.py`` to
two pytest functions matching the slice contract:
- test_no_get_torch_device_in_forward_methods: AST-walks every
``forward`` method body in comfy/ldm/seedvr/model.py and asserts
none calls ``get_torch_device``. Reports ``(func_lineno,
call_lineno)`` so the pre-fix failure produces ``[(613, 624)]``.
- test_mmmodule_forward_succeeds_without_get_torch_device_lookup:
monkeypatches ``comfy.model_management.get_torch_device`` to raise,
then runs ``MMModule(nn.Linear, 16, 8, shared_weights=False)``
forward. With the cast removed the call must complete; with the
cast present it raises the planted RuntimeError.
Closes pollockjj/mydevelopment#185
… slice 1 contract Rewrite the second test to match the contract shape: zero-arg boom that increments call_count[0] before raising, local _IdentityCallable nn.Module returning x, MMModule(_IdentityCallable, shared_weights=False, vid_only=False), mm.forward(torch.zeros(2, 4), torch.ones(2, 4)), and assert call_count[0] == 0 plus identity/device equality on both outputs. Rewrite the first test to use the contract list-comprehension shape against ast.parse(inspect.getsource(...)) on the seedvr model module.
Codex Review — Cycle 1Overall: patch is correct No findings. |
There was a problem hiding this comment.
Pull request overview
Removes the last SeedVR2 MMModule.forward device lookup/cast that bypassed ComfyUI's model-management placement, and adds a regression test module to lock that invariant in place.
Changes:
- Deletes the
comfy.model_management.get_torch_device()lookup andvid.to(device)cast fromMMModule.forward. - Adds an AST-based regression test asserting no
forwardmethod incomfy.ldm.seedvr.modelcallsget_torch_device. - Adds a smoke test proving
MMModule.forwardcompletes without touchingget_torch_device.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.
| File | Description |
|---|---|
comfy/ldm/seedvr/model.py |
Removes the residual forward-path global device cast from MMModule.forward. |
tests-unit/comfy_test/test_seedvr_forward_no_device_cast.py |
Adds regression coverage for the no-lookup invariant and a direct MMModule.forward smoke test. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Melian PR-Review — Cycle 1 — DECISION: STOP — PR-review cycle converged.HEAD SHA reviewed: Why we stoppedPer the PR-review gate rule (2026-05-05), the cycle stops iff both hold:
Both conditions are met. Any remaining advisory findings below are on test/doc paths or have been ACCEPT-classified by the slicer — they are converged work, not blockers. Raw signal
Verdict rationale (raw)
This comment is posted by Melian's FSM ( |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Plan: Remove residual
comfy.model_management.get_torch_device()cast fromMMModule.forwardOverview
comfy/ldm/seedvr/model.pyonpollockjj/ComfyUI:origin/issue_101HEAD98df2309carries one residual forward-path device lookup that fights Comfy's model-management system:MMModule.forward(file lines 613–631) readsdevice = comfy.model_management.get_torch_device()at line 624 and forces the input viavid = vid.to(device)at line 625 — the exact pattern Kosinkadink flagged on Comfy-Org#11294 (issue-comment4259303182, item 4) and the only one of Kosinkadink's five flagged sites that survived Yousef'sFixed 3, 4reply. Phase 0 ground-truth probes (GTP-2, GTP-3) confirm oneget_torch_device()call in anyforwardmethod body across the file (MMModule.forward, line 624). The slice removes those two lines, locks the repository invariant via an AST-walked regression that asserts zeroget_torch_device()calls inside anyforwardmethod body in this file, and adds a behavioral smoke test that monkey-patchescomfy.model_management.get_torch_deviceto raise — proving the post-fix forward never looks it up — while exercisingMMModule.forward(vid, txt)end-to-end against annn.Identity-shaped stub on CPU. Companion GTP-3 reports one.to(device=..., dtype=...)keyword cast at line 629; that cast is a derived-device cast (device=vid.device) sourced from the input stream, not from a Comfy lookup, and is left untouched per the issue body's specific scope to "lines 624-625" (Out of Scope §). The fix lands onpollockjj/ComfyUI:issue_185cut fromorigin/issue_101HEAD98df2309and PRs back intoissue_101, mirroring the merged sequence #46 / #44 / #43 / #36 / #35 / #34 / #33 / #32 / #31.Diagnosis Summary
The buggy method body, captured live from
origin/issue_101:comfy/ldm/seedvr/model.pylines 613–631 (Ground Truth Probe GTP-4):Lines 624 (
device = comfy.model_management.get_torch_device()) and 625 (vid = vid.to(device)) are the residual cast. Comfy's model-management system places weights and inputs onload_devicebeforeforwardis invoked; an inner module re-querying the globalget_torch_device()and forcing the input there bypasses Comfy's per-modelload_devicedecision (which can differ from the global probe under partial offload, low-VRAM, or CPU-fallback configurations). Removing both lines letsvidride whatever device Comfy passed in, which is the Comfy-conventional contract.The method-local
txt = txt.to(device=vid.device, dtype=vid.dtype)at line 629 is structurally distinct: its device target isvid.device(derived from the input stream, post-vid_modulecall), notcomfy.model_management.get_torch_device(). Phase 0 reports it for completeness but the issue body explicitly scopes the deliverable to "the remaining cast atMMModule.forward(lines 624-625)" — Kosinkadink'sget_torch_device()+.to(device)pattern. Line 629 also carries adtype=vid.dtypecomponent whose load-bearing role for cross-stream dtype alignment is not in scope for this audit (Out of Scope §).Affected Repositories
98df2309)pollockjj/ComfyUIis the deliverable: the edit tocomfy/ldm/seedvr/model.py(MMModule.forward, lines 624–625 removed) and the new regression test undertests-unit/comfy_test/test_seedvr_forward_no_device_cast.pyland onissue_185cut fromorigin/issue_101HEAD98df2309and PR back intoissue_101. This mirrors the merged eval-and-fix PR sequence #46 (Comfy-Org#192), #44 (Comfy-Org#190), #43 (Comfy-Org#194), #36 (Comfy-Org#119), #35 (Comfy-Org#188), #34 (Comfy-Org#189), #33 (#109), #32 (Comfy-Org#183), #31 (groupnorm).pollockjj/mydevelopmentis the bookkeeping target: every committed artifact undergithub_issues/185/slice1/(Phase 0 probe transcript, A/B pytest logs, AST-extracted forward source, scope-bounded diff, test source copy) lives on the outer workspace'sissue_185branch and PRs back intomainas the standard[bookkeeping] [Issue 101][eval-and-fix]pattern.Asset Readiness
git rev-parse origin/issue_101=98df23092.11.0+cu130, CUDA available98df2309)comfy.model_management.get_torch_deviceimport targetMMModule/MMArg/get_args/get_kwargssymbol contractsgithub_issues/185/slice1/writable onissue_185Research and Methodology
Plan Foundations Comment URL: https://github.com/pollockjj/mydevelopment/issues/185#issuecomment-4383764590 — Foundations comment for Comfy-Org#185 is the load-bearing reference for the inline R&M and TPM content reproduced verbatim below; every AC carrying a
Foundations-pin:line cites the comment node-id andcreatedAttimestamp (lastEditedAtis null on this comment, so qa-plan Check 21 sub-check 6 falls back tocreatedAtas the authoritative pin).Detected Scope:
none— port of the established eval-and-fix pattern to one residual cast site flagged by Kosinkadink on the upstream PR review (Comfy-Org#11294 issue-comment4259303182, item 4); existing Comfy model-management contract IS the equivalence reference, not a published metric. Equivalence rule is binary source-AST equality (zeroget_torch_device()Call nodes inside anyforwardFunctionDef body incomfy/ldm/seedvr/model.py) plus binary behavioral equality (post-fixMMModule.forward(vid, txt)runs to completion withcomfy.model_management.get_torch_devicemonkey-patched to raise, returning tensors on the same device as the inputs). No tolerance band, no statistical test, no measurement instrument.The change class is "honor an existing internal API contract" — Comfy's per-model
load_deviceplacement contract forforwardinvocations. Source authority is the live source ofcomfy/ldm/seedvr/model.pyon the issue's base branch (origin/issue_101HEAD98df2309), captured verbatim under## Ground Truth Probes. Binding parent-issue directives quoted verbatim from pollockjj/mydevelopment#185 issue body:These directives bound every AC below. Diff scope is confined to
MMModule.forwardbody (lines 613–631 on theorigin/issue_101side); the parent class structure-resolution helper and RoPE cache touched by #110 are off-limits; upstream PR Comfy-Org#11294 is not perturbed.Tools, Pipeline, and Measurements
Plan Foundations Comment URL: https://github.com/pollockjj/mydevelopment/issues/185#issuecomment-4383764590 — same Foundations comment as the R&M section above; the TPM content reproduced verbatim below is the load-bearing reference for tool / version / pipeline citations in every AC carrying a
Foundations-pin:line.Trivial scope; no external tools beyond the change itself.
pytesttests-unit/requirements.txtpin (pertests-unit/README.md)test_seedvr_groupnorm_limit.py,seedvr_model_test.pyexercise the same venv onorigin/issue_101torch2.11.0+cu130on/home/johnj/dev_cuda_1/ComfyUI/.venv/comfy.model_management.get_torch_devicecomfy/model_management.pyimport path on slot venvcomfy.ldm.seedvr.model.MMModule/MMArg/get_args/get_kwargsgitast,inspect,re,unittest.mock(viamonkeypatchfixture).venv/seedvr_model_test.pyAST pattern)Pipeline: edit one method body in one source file (
comfy/ldm/seedvr/model.py,MMModule.forward, lines 624–625 on theorigin/issue_101side, removing the two lines verbatim) → add one new test filetests-unit/comfy_test/test_seedvr_forward_no_device_cast.pydefining two pytest test functions → runpytestagainst the new test onissue_185HEAD (post-fix capture) and against the same test withcomfy/ldm/seedvr/model.pyreverted toorigin/issue_101(pre-fix capture) → commit pytest logs, AST-extracted forward source, grep transcripts, and scope-boundedgit difftogithub_issues/185/slice1/on the bookkeeping branch.Measurements: zero numeric measurements. The two evidence primitives are (a) source-AST count of
get_torch_device()Call nodes insideforwardFunctionDef bodies (binary, must equal0post-fix) and (b) behavioral pytest pass/fail on a forward invocation withget_torch_devicemonkey-patched to raise (binary, must PASS post-fix). HR-06 (single-probe-before-sweep) and HR-11 (boundary-bracketing) do not apply — no sweep, no metric distribution.Ground Truth Probes
Every literal API surface the ACs below cite is anchored to a probe captured live against the issue's read target (
pollockjj/ComfyUI:origin/issue_101HEAD98df2309) or against the active PyTorch interpreter at plan-authoring time (/home/johnj/dev_cuda_1/ComfyUI/.venv/bin/python, torch2.11.0+cu130, CUDA available).cd /home/johnj/dev_cuda_1/ComfyUI && git show origin/issue_101:comfy/ldm/seedvr/model.py | python3 -c 'import ast,sys; tree=ast.parse(sys.stdin.read()); [print(f"forward {n.lineno}-{n.end_lineno}") for n in ast.walk(tree) if isinstance(n,ast.FunctionDef) and n.name=="forward"]'forward 34-51forward 369-399forward 422-439forward 511-548forward 613-631forward 686-687forward 731-828forward 842-846forward 864-865forward 932-969forward 984-993forward 996-1018forward 1033-1043forward 1046-1064forward 1095-1145forward 1167-1189forward 1387-1471— 17forwardFunctionDefs total; theMMModule.forwardbody the AC operates on spans lines 613–631.cd /home/johnj/dev_cuda_1/ComfyUI && git show origin/issue_101:comfy/ldm/seedvr/model.py | python3 -c 'import ast,sys; t=ast.parse(sys.stdin.read()); cls={id(m):c.name for c in ast.walk(t) if isinstance(c,ast.ClassDef) for m in c.body if isinstance(m,ast.FunctionDef)}; [print(f"class={cls.get(id(n))} forward.lineno={n.lineno} call.lineno={i.lineno}") for n in ast.walk(t) if isinstance(n,ast.FunctionDef) and n.name=="forward" for i in ast.walk(n) if isinstance(i,ast.Call) and isinstance(i.func,ast.Attribute) and i.func.attr=="get_torch_device"]'class=MMModule forward.lineno=613 call.lineno=624— exactly oneget_torch_device()Call node inside anyforwardmethod body in the entire file, located insideMMModule.forwardat line 624. The post-fix expectation is zero matches.cd /home/johnj/dev_cuda_1/ComfyUI && git show origin/issue_101:comfy/ldm/seedvr/model.py | python3 -c 'import ast,sys; t=ast.parse(sys.stdin.read()); cls={id(m):c.name for c in ast.walk(t) if isinstance(c,ast.ClassDef) for m in c.body if isinstance(m,ast.FunctionDef)}; [print(f"class={cls.get(id(n))} forward.lineno={n.lineno} call.lineno={i.lineno}") for n in ast.walk(t) if isinstance(n,ast.FunctionDef) and n.name=="forward" for i in ast.walk(n) if isinstance(i,ast.Call) and isinstance(i.func,ast.Attribute) and i.func.attr=="to" and any(kw.arg=="device" for kw in i.keywords)]'class=MMModule forward.lineno=613 call.lineno=629— exactly one.to(device=...)keyword cast inside anyforwardmethod body, located at MMModule.forward line 629 (txt = txt.to(device=vid.device, dtype=vid.dtype)). Its device argument derives fromvid.device(input-stream device), NOT fromcomfy.model_management.get_torch_device(); this is reported by the Phase 0 probe as a structurally distinct pattern from the lines 624-625 deliverable scope (Out of Scope §).cd /home/johnj/dev_cuda_1/ComfyUI && git show origin/issue_101:comfy/ldm/seedvr/model.py | sed -n '613,631p'device = comfy.model_management.get_torch_device()) and 625 (vid = vid.to(device)); every other line of the body is preserved verbatim.cd /home/johnj/dev_cuda_1/ComfyUI && git ls-tree --name-only origin/issue_101 tests-unit/comfy_test/ tests-unit/comfy_extras_test/ | grep -E 'seedvr'origin/issue_101HEAD; AC-6 enumerates them all and asserts each remains green onissue_185HEAD. The new filetest_seedvr_forward_no_device_cast.pydoes not yet exist.cd /home/johnj/dev_cuda_1/ComfyUI && git show origin/issue_101:comfy/ldm/seedvr/model.py | python3 -c 'import ast,sys; t=ast.parse(sys.stdin.read()); want={"MMArg","get_args","get_kwargs","MMModule"}; [print(f"{n.name}: kind={type(n).__name__} lineno={n.lineno} end={n.end_lineno}") for n in ast.walk(t) if (isinstance(n,ast.ClassDef) or isinstance(n,ast.FunctionDef)) and n.name in want]'MMArg: kind=ClassDef lineno=154 end=156get_args: kind=FunctionDef lineno=177 end=178get_kwargs: kind=FunctionDef lineno=181 end=182MMModule: kind=ClassDef lineno=589 end=631— the four symbols the smoke test scaffold imports/instantiates exist at the cited line ranges;MMModule.__init__acceptsmodule: Callable[..., nn.Module], *args, shared_weights: bool=False, vid_only: bool=False, **kwargsand instantiatesself.vid/self.txtviamodule(*get_args("vid", args), **get_kwargs("vid", kwargs)).cd /home/johnj/dev_cuda_1/ComfyUI && /home/johnj/dev_cuda_1/ComfyUI/.venv/bin/python -c 'import comfy.model_management as m, inspect; print("callable=",callable(m.get_torch_device)); print("sig=",inspect.signature(m.get_torch_device))'callable= Truesig= ()—comfy.model_management.get_torch_deviceis a zero-arg callable on the slot venv. The smoke test monkey-patches it viamonkeypatch.setattr(comfy.model_management, "get_torch_device", boom)whereboomis a zero-arg function that increments a counter and raisesRuntimeError; pre-fix, thedevice = comfy.model_management.get_torch_device()lookup at MMModule.forward line 624 invokesboom, the counter increments, andRuntimeErrorpropagates out offorward. Post-fix, the lookup is gone and the counter remains zero.cd /home/johnj/dev_cuda_1/ComfyUI && /home/johnj/dev_cuda_1/ComfyUI/.venv/bin/python -c 'import torch; print("torch", torch.__version__); print("cuda_available", torch.cuda.is_available())'torch 2.11.0+cu130cuda_available True— slot venv is CUDA-capable; on a CUDA-host run, an unpatchedget_torch_device()returns a CUDA device, so the pre-fixvid.to(device)cast at line 625 actually moves CPU input tensors to GPU. The smoke test pins inputs totorch.device("cpu")and the monkey-patchedget_torch_deviceraises before any device move can occur, so the pre-fix failure mode is theRuntimeErrorfromboom(not a CUDA-vs-CPU device-mismatch downstream).If any literal in an AC is not anchored above, the AC is REJECTED. The slicer must replicate the exact GTP commands (byte-identical) when assembling the slice — every cited file:line, function name, class name, and Python literal in the slice's commit must match the GTP output above.
Created Surface Contract
The plan introduces the following literals; they are anchored by the slice that creates them, not by Phase 0 probes (these strings do not exist on the read target — the slicer writes them):
tests-unit/comfy_test/test_seedvr_forward_no_device_cast.pypollockjj/ComfyUIonissue_185. Defines exactly two pytest test functions: (1)test_no_get_torch_device_in_forward_methods— usesinspect.getsource(comfy.ldm.seedvr.model)thenast.parse(...)and asserts that the list[(n.lineno, i.lineno) for n in ast.walk(tree) if isinstance(n, ast.FunctionDef) and n.name == "forward" for i in ast.walk(n) if isinstance(i, ast.Call) and isinstance(i.func, ast.Attribute) and i.func.attr == "get_torch_device"]is empty. (2)test_mmmodule_forward_succeeds_without_get_torch_device_lookup— uses themonkeypatchfixture to replacecomfy.model_management.get_torch_devicewith a zero-arg callable that incrementscall_count[0]and raisesRuntimeError("MMModule.forward called get_torch_device()"); constructsMMModule(_IdentityCallable, shared_weights=False, vid_only=False)where_IdentityCallableis a localnn.Modulesubclass whoseforward(x, *args, **kwargs)returnsx; callsmm.forward(torch.zeros(2, 4), torch.ones(2, 4)); asserts (a)call_count[0] == 0, (b)torch.equal(vid_out, vid_in), (c)torch.equal(txt_out, txt_in), (d)vid_out.device == vid_in.device, (e)txt_out.device == txt_in.device. The test setscomfy.cli_args.args.cpu = Truewhennot torch.cuda.is_available()(matches thetest_seedvr_groupnorm_limit.pyimport preamble).github_issues/185/slice1/phase0_probe.logpollockjj/mydevelopmentissue_185. Verbatim stdout of the two AST probes (GTP-2 and GTP-3) captured againstorigin/issue_101HEAD98df2309before any slice edits. Final file content matches GTP-2'sclass=MMModule forward.lineno=613 call.lineno=624line and GTP-3'sclass=MMModule forward.lineno=613 call.lineno=629line, with a header recording the read-target SHA.github_issues/185/slice1/pytest_post_fix.logpollockjj/mydevelopmentissue_185. Verbatim stdout/stderr ofcd /home/johnj/dev_cuda_1/ComfyUI && python3 -m pytest tests-unit/comfy_test/test_seedvr_forward_no_device_cast.py -vrun onissue_185HEAD. Final non-blank summary line ends with2 passed.github_issues/185/slice1/pytest_pre_fix.logpollockjj/mydevelopmentissue_185. Verbatim stdout/stderr of the same pytest invocation captured aftercd /home/johnj/dev_cuda_1/ComfyUI && git checkout origin/issue_101 -- comfy/ldm/seedvr/model.pyis applied to the deliverable working tree (the new test file fromissue_185HEAD remains in place; onlymodel.pyis reverted). Captures oneassert not foundfailure ontest_no_get_torch_device_in_forward_methodsreporting[(613, 624)]plus oneRuntimeError: MMModule.forward called get_torch_device()failure ontest_mmmodule_forward_succeeds_without_get_torch_device_lookup. The deliverable working tree is restored toissue_185HEAD viacd /home/johnj/dev_cuda_1/ComfyUI && git restore --source=HEAD comfy/ldm/seedvr/model.pyimmediately after capture, before any commit on the deliverable.github_issues/185/slice1/forward_source.txtpollockjj/mydevelopmentissue_185. Verbatim AST-extracted source ofMMModule.forwardfromissue_185HEADcomfy/ldm/seedvr/model.py, produced bycd /home/johnj/dev_cuda_1/ComfyUI && python3 -c 'import ast; src=open("comfy/ldm/seedvr/model.py").read(); tree=ast.parse(src); cls=[n for n in ast.walk(tree) if isinstance(n,ast.ClassDef) and n.name=="MMModule"][0]; m=[n for n in cls.body if isinstance(n,ast.FunctionDef) and n.name=="forward"][0]; print(ast.get_source_segment(src,m))'. The file content shows zero occurrences of the substringscomfy.model_management.get_torch_deviceandvid.to(device).github_issues/185/slice1/forbidden_call_grep.logpollockjj/mydevelopmentissue_185. Output ofgrep -n -E 'comfy\.model_management\.get_torch_device|vid = vid\.to\(device\)' github_issues/185/slice1/forward_source.txt; echo "exit=$?". The committed log recordsexit=1and zero matching lines.github_issues/185/slice1/scope_diff.txtpollockjj/mydevelopmentissue_185. Output ofcd /home/johnj/dev_cuda_1/ComfyUI && git diff origin/issue_101..issue_185 -- comfy/ldm/seedvr/model.py. The diff shows two deleted lines (the verbatim text of original lines 624 and 625 from GTP-4) and zero hunks outside the body ofMMModule.forward(lines 613–631 on theorigin/issue_101side). The artifact contains the verbatim diff plus a final stanza listing every@@-hunk header range and confirming each falls inside[613, 631]on the-side.github_issues/185/slice1/test_source.txtpollockjj/mydevelopmentissue_185. Verbatim copy produced bycp /home/johnj/dev_cuda_1/ComfyUI/tests-unit/comfy_test/test_seedvr_forward_no_device_cast.py github_issues/185/slice1/test_source.txt. AC-5 verifies the file defines exactly the two named test functions, the AST assertion shape, and the monkeypatch boom-counter assertion shape.github_issues/185/slice1/companion_pytest.logpollockjj/mydevelopmentissue_185. Verbatim stdout/stderr ofcd /home/johnj/dev_cuda_1/ComfyUI && python3 -m pytest tests-unit/comfy_test/seedvr_model_test.py tests-unit/comfy_test/seedvr_vae_forward_test.py tests-unit/comfy_test/seedvr_vae_wrapper_forward_test.py tests-unit/comfy_test/test_seedvr_groupnorm_limit.py tests-unit/comfy_test/test_seedvr_rope_delegation.py tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.py tests-unit/comfy_test/test_seedvr_vae_decode_guards.py tests-unit/comfy_test/test_seedvr_vae_decode_unpadded_t.py tests-unit/comfy_test/test_seedvr_vae_loader_metadata.py tests-unit/comfy_test/test_seedvr_vae_tiled_args_no_mutate.py tests-unit/comfy_extras_test/test_seedvr_conditioning_hardening.py tests-unit/comfy_extras_test/test_seedvr_node_signature.py -vrun onissue_185HEAD. Final non-blank summary line ends withpassedand zerofailed/ zeroerrors. The 12 file paths enumerated in the command match GTP-5 verbatim.Slices
Slice 1: Remove residual
get_torch_device()cast fromMMModule.forward+ lock invariant + smoke testKind: implementation
Objective: Prove that
MMModule.forward()no longer readscomfy.model_management.get_torch_device()nor forcesvid = vid.to(device), that the change is scoped strictly to the body ofMMModule.forward(no other method body, no other class touched), that the repository-level invariant againstget_torch_device()calls inside anyforwardmethod body incomfy/ldm/seedvr/model.pyis locked in by an AST-walked regression, that annn.Identity-shaped smoke throughMMModule.forward(vid, txt)succeeds without invokingget_torch_device(verified viamonkeypatchthat raises if invoked) and preserves input device placement, and that the existing 12-test SeedVR2 surface remains pytest-green onissue_185HEAD after the edit lands.Acceptance Criteria
cd /home/johnj/dev_cuda_1/ComfyUI && python3 -m pytest tests-unit/comfy_test/test_seedvr_forward_no_device_cast.py -vexits 0 with2 passedin the summary line — verified by committed artifactgithub_issues/185/slice1/pytest_post_fix.logwhose last non-blank line ends with2 passed.origin/issue_101source raisesassert not foundontest_no_get_torch_device_in_forward_methods(reporting[(613, 624)]per GTP-2) andRuntimeError: MMModule.forward called get_torch_device()ontest_mmmodule_forward_succeeds_without_get_torch_device_lookup(per GTP-4 line 624 and GTP-7'sboomsemantics)2 passedMMModule.forward()no longer readscomfy.model_management.get_torch_device()nor forcesvid = vid.to(device), that the change is scoped strictly to the body ofMMModule.forward(no other method body, no other class touched), that the repository-level invariant againstget_torch_device()calls inside anyforwardmethod body incomfy/ldm/seedvr/model.pyis locked in by an AST-walked regression, that annn.Identity-shaped smoke throughMMModule.forward(vid, txt)succeeds without invokingget_torch_device(verified viamonkeypatchthat raises if invoked) and preserves input device placement, and that the existing 12-test SeedVR2 surface remains pytest-green onissue_185HEAD after the edit lands."tests-unit/comfy_test/test_seedvr_forward_no_device_cast.py -vproves the same observables flip across the fix:github_issues/185/slice1/pytest_pre_fix.logrecords theassert not foundfailure ontest_no_get_torch_device_in_forward_methods(reporting[(613, 624)]) plus theRuntimeError: MMModule.forward called get_torch_device()failure ontest_mmmodule_forward_succeeds_without_get_torch_device_lookupaftercd /home/johnj/dev_cuda_1/ComfyUI && git checkout origin/issue_101 -- comfy/ldm/seedvr/model.py, andgithub_issues/185/slice1/pytest_post_fix.logrecords pytest exit 0 with2 passedaftercd /home/johnj/dev_cuda_1/ComfyUI && git restore --source=HEAD comfy/ldm/seedvr/model.pyrestoresissue_185HEAD.comfy.model_management.get_torch_deviceis the zero-arg callable that the monkey-patchedboomreplaces.issue_185HEAD exits 0 with2 passedand zeroassert not found/RuntimeErrorfailures.MMModule.forward()no longer readscomfy.model_management.get_torch_device()nor forcesvid = vid.to(device), that the change is scoped strictly to the body ofMMModule.forward(no other method body, no other class touched), that the repository-level invariant againstget_torch_device()calls inside anyforwardmethod body incomfy/ldm/seedvr/model.pyis locked in by an AST-walked regression, that annn.Identity-shaped smoke throughMMModule.forward(vid, txt)succeeds without invokingget_torch_device(verified viamonkeypatchthat raises if invoked) and preserves input device placement, and that the existing 12-test SeedVR2 surface remains pytest-green onissue_185HEAD after the edit lands."MMModule.forwardonissue_185HEAD contains zero occurrences of the substringscomfy.model_management.get_torch_deviceandvid = vid.to(device)— verified by committed artifactsgithub_issues/185/slice1/forward_source.txt(the extracted body) andgithub_issues/185/slice1/forbidden_call_grep.log(grep -n -E '...' forward_source.txt; echo "exit=$?"recordsexit=1, zero matching lines). The remaining body lines from GTP-4 (the four-linedef forward(...)header through theTuple[...]return type, thevid_module = ...line, thevid = vid_module(vid, ...)line, theif not self.vid_only:block including line 629'stxt = txt.to(device=vid.device, dtype=vid.dtype)per GTP-3, and the finalreturn vid, txt) are preserved verbatim. Note: GTP-3's.to(device=vid.device, dtype=vid.dtype)cast at line 629 is not removed by this slice; it is a derived-device cast (Out of Scope §) and is reported by Phase 0 for completeness only.comfy.model_management.get_torch_deviceat line 624 andvid = vid.to(device)at line 625 of the pre-fix forward body.github_issues/185/slice1/forward_source.txtshows neither substring;github_issues/185/slice1/forbidden_call_grep.logrecords exit code 1 with zero matches.MMModule.forward()no longer readscomfy.model_management.get_torch_device()nor forcesvid = vid.to(device), that the change is scoped strictly to the body ofMMModule.forward(no other method body, no other class touched), that the repository-level invariant againstget_torch_device()calls inside anyforwardmethod body incomfy/ldm/seedvr/model.pyis locked in by an AST-walked regression, that annn.Identity-shaped smoke throughMMModule.forward(vid, txt)succeeds without invokingget_torch_device(verified viamonkeypatchthat raises if invoked) and preserves input device placement, and that the existing 12-test SeedVR2 surface remains pytest-green onissue_185HEAD after the edit lands."cd /home/johnj/dev_cuda_1/ComfyUI && git diff origin/issue_101..issue_185 -- comfy/ldm/seedvr/model.pyproduces hunks confined to the line range ofMMModule.forwardon theorigin/issue_101side (lines 613–631 per GTP-1) and zero hunks touching any of the other 16forwardmethods enumerated by GTP-1 (forward 34-51,369-399,422-439,511-548,686-687,731-828,842-846,864-865,932-969,984-993,996-1018,1033-1043,1046-1064,1095-1145,1167-1189,1387-1471) or any other class definition in the file — verified by committed artifactgithub_issues/185/slice1/scope_diff.txt. The diff shows exactly two deleted lines (text matching original lines 624 and 625 from GTP-4) and zero added lines. The artifact contains the verbatim diff plus a final stanza listing every@@-hunk header range and confirming each falls inside[613, 631]on the-side.forwardmethods; GTP-2 localizes the onlyget_torch_device()call site to MMModule.forward line 624. Any diff hunk outside[613, 631]is extraneous to the diagnosed defect.github_issues/185/slice1/scope_diff.txtenumerates only in-range hunks with two deletions and zero additions.MMModule.forward()no longer readscomfy.model_management.get_torch_device()nor forcesvid = vid.to(device), that the change is scoped strictly to the body ofMMModule.forward(no other method body, no other class touched), that the repository-level invariant againstget_torch_device()calls inside anyforwardmethod body incomfy/ldm/seedvr/model.pyis locked in by an AST-walked regression, that annn.Identity-shaped smoke throughMMModule.forward(vid, txt)succeeds without invokingget_torch_device(verified viamonkeypatchthat raises if invoked) and preserves input device placement, and that the existing 12-test SeedVR2 surface remains pytest-green onissue_185HEAD after the edit lands."tests-unit/comfy_test/test_seedvr_forward_no_device_cast.pyonissue_185defines exactly two pytest test functions whose names aretest_no_get_torch_device_in_forward_methodsandtest_mmmodule_forward_succeeds_without_get_torch_device_lookup. The first runstree = ast.parse(inspect.getsource(comfy.ldm.seedvr.model))and asserts[(n.lineno, i.lineno) for n in ast.walk(tree) if isinstance(n, ast.FunctionDef) and n.name == "forward" for i in ast.walk(n) if isinstance(i, ast.Call) and isinstance(i.func, ast.Attribute) and i.func.attr == "get_torch_device"] == []. The second usesmonkeypatch.setattr(comfy.model_management, "get_torch_device", boom)whereboomis a zero-arg callable that incrementscall_count[0]and raisesRuntimeError, instantiatesMMModule(_IdentityCallable, shared_weights=False, vid_only=False)with a local_IdentityCallable(nn.Module)whoseforward(x, *args, **kwargs)returnsx, invokesmm.forward(torch.zeros(2, 4), torch.ones(2, 4)), and asserts (a)call_count[0] == 0, (b)torch.equal(vid_out, vid_in), (c)torch.equal(txt_out, txt_in), (d)vid_out.device == vid_in.device, (e)txt_out.device == txt_in.device— verified by committed artifactgithub_issues/185/slice1/test_source.txt(cp /home/johnj/dev_cuda_1/ComfyUI/tests-unit/comfy_test/test_seedvr_forward_no_device_cast.py github_issues/185/slice1/test_source.txt) plus the AC-1 pytest log showing both test names in thePASSEDlines.test_seedvr_forward_no_device_cast.pydoes not exist onorigin/issue_101HEAD.github_issues/185/slice1/test_source.txtadds the two named tests andgithub_issues/185/slice1/pytest_post_fix.logshows both PASS.MMModule.forward()no longer readscomfy.model_management.get_torch_device()nor forcesvid = vid.to(device), that the change is scoped strictly to the body ofMMModule.forward(no other method body, no other class touched), that the repository-level invariant againstget_torch_device()calls inside anyforwardmethod body incomfy/ldm/seedvr/model.pyis locked in by an AST-walked regression, that annn.Identity-shaped smoke throughMMModule.forward(vid, txt)succeeds without invokingget_torch_device(verified viamonkeypatchthat raises if invoked) and preserves input device placement, and that the existing 12-test SeedVR2 surface remains pytest-green onissue_185HEAD after the edit lands."cd /home/johnj/dev_cuda_1/ComfyUI && python3 -m pytest tests-unit/comfy_test/seedvr_model_test.py tests-unit/comfy_test/seedvr_vae_forward_test.py tests-unit/comfy_test/seedvr_vae_wrapper_forward_test.py tests-unit/comfy_test/test_seedvr_groupnorm_limit.py tests-unit/comfy_test/test_seedvr_rope_delegation.py tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.py tests-unit/comfy_test/test_seedvr_vae_decode_guards.py tests-unit/comfy_test/test_seedvr_vae_decode_unpadded_t.py tests-unit/comfy_test/test_seedvr_vae_loader_metadata.py tests-unit/comfy_test/test_seedvr_vae_tiled_args_no_mutate.py tests-unit/comfy_extras_test/test_seedvr_conditioning_hardening.py tests-unit/comfy_extras_test/test_seedvr_node_signature.py -vexits 0 onissue_185HEAD with the final summary line containingpassedand zerofailed/ zeroerrors— verified by committed artifactgithub_issues/185/slice1/companion_pytest.log. The 12 file paths in the command match GTP-5 verbatim.github_issues/185/slice1/pytest_pre_fix.logrecords the dedicated failing behavior under audit on the pre-fix tree:assert not foundwith[(613, 624)]plusRuntimeError: MMModule.forward called get_torch_device(), while the 12 legacy SeedVR2 tests listed in GTP-5 are already green onorigin/issue_101.github_issues/185/slice1/pytest_post_fix.logflips the dedicated new test to2 passed, andgithub_issues/185/slice1/companion_pytest.logshows the 12 legacy SeedVR2 tests remain green with zerofailed/ zeroerrors.MMModule.forward()no longer readscomfy.model_management.get_torch_device()nor forcesvid = vid.to(device), that the change is scoped strictly to the body ofMMModule.forward(no other method body, no other class touched), that the repository-level invariant againstget_torch_device()calls inside anyforwardmethod body incomfy/ldm/seedvr/model.pyis locked in by an AST-walked regression, that annn.Identity-shaped smoke throughMMModule.forward(vid, txt)succeeds without invokingget_torch_device(verified viamonkeypatchthat raises if invoked) and preserves input device placement, and that the existing 12-test SeedVR2 surface remains pytest-green onissue_185HEAD after the edit lands."Constraints
python3python3 -m pytest <test-path> -vinvoked from the deliverable checkout root/home/johnj/dev_cuda_1/ComfyUI.venv/already providestorch 2.11.0+cu130andpytestper GTP-8 and the slot layout in CLAUDE.mdpkill, norm -rf, nopython main.py, no/tmpwrites&&/||/;chainingpollockjj/ComfyUI:issue_185cut fromorigin/issue_101(HEAD98df2309); PR target =issue_101github_issues/185/slice1/*) land onpollockjj/mydevelopment:issue_185cut frommain; PR target =main(standard[bookkeeping] [Issue 101][eval-and-fix]pattern)MMModule.forwardincomfy/ldm/seedvr/model.py; the other 16forwardmethods enumerated by GTP-1 are off-limits, as areMMModule.__init__, every other class in the file (CustomRMSNorm,MMArg,RotaryEmbeddingBase,MMRotaryEmbeddingBase,NaMMRotaryEmbedding3d,NaMMAttention,AdaLayerNorm,TimeEmbedding,NaDiT, etc.), and every module-level function (safe_pad_operation,get_args,get_kwargs,get_window_op, etc.)comfy/ldm/seedvr/vae.py,comfy_extras/nodes_seedvr.py,comfy/sd.py, or any other file in the deliverable repoMMModule.forward's signature or return-type annotation (thedef forward(self, vid, txt, *args, **kwargs) -> Tuple[torch.FloatTensor, torch.FloatTensor]:shape from GTP-4 lines 613–622 is preserved verbatim)Comfy-Org/ComfyUI#11294oryousef-rafat:seedvr2pytestfrom the deliverable slot's.venv/— REUSE; no installcomfy.ldm.seedvr.model.MMModule/MMArg/get_args/get_kwargsimport paths — REUSE; verified by GTP-6comfy.model_management.get_torch_device— REUSE; verified by GTP-7comfy.cli_args.args.cpu = Trueimport-time guard pattern — REUSE; established intests-unit/comfy_test/test_seedvr_groupnorm_limit.pyandtests-unit/comfy_test/seedvr_model_test.pygit,grep, AST extraction via stdlibastmodule,monkeypatchfixture — REUSEgithub_issues/185/slice1/are the only added paths in the bookkeeping repoMMModule.forwardstops callingget_torch_device()) → A/B pytest pair: post-fix log onissue_185HEAD plus pre-fix log captured withmodel.pytransiently reverted toorigin/issue_101(Slice 1 AC-1, AC-2)comfy.model_management.get_torch_deviceand novid = vid.to(device)in MMModule.forward body) → AST-extracted committed source plus zero-match grep transcript (Slice 1 AC-3)git diff origin/issue_101..issue_185 -- comfy/ldm/seedvr/model.pywith hunk-range enumeration (Slice 1 AC-4)git stashor external decision deferral..gitignorerules for safe generated bulk, or discarded generated debris.issue_185(e.g. the?? github_issues/185/dispatch artifacts seen at plan time on the bookkeeping checkout) is repo provenance and is integrated, not removed as dirty-state cleanup.git status --shortgates locally on both/home/johnj/dev_cuda_1/ComfyUIand/home/johnj/dev_cuda_1/mydevelopmentand must refuse to submit while either is dirty.git checkout origin/issue_101 -- comfy/ldm/seedvr/model.py) MUST be undone viagit restore --source=HEAD comfy/ldm/seedvr/model.pybefore any commit is created on the deliverable; the pre-fix log itself is committed only to the bookkeeping repo.Out of Scope
comfy/ldm/seedvr/model.py:MMModule.forward:629txt = txt.to(device=vid.device, dtype=vid.dtype)— Phase 0 GTP-3 reports this as the only.to(device=...)keyword-form cast inside anyforwardmethod body in the file. Its device argument isvid.device(input-stream-derived), notcomfy.model_management.get_torch_device(); it is structurally distinct from Kosinkadink's flagged pattern (get_torch_device()lookup →.to(device)cast on the looked-up value). The issue body explicitly scopes the deliverable to "the remaining cast atMMModule.forward(lines 624-625)" — GTP-2's single match — not to every.to(device=...)call site. Line 629 also carries adtype=vid.dtypecomponent whose load-bearing role for cross-stream dtype alignment is a separate audit question that this plan does not adjudicate. Out of scope for this issue; if a follow-up audit determines line 629 is also a violation, file a separate issue.forwardmethods enumerated by GTP-1 (forward 34-51,369-399,422-439,511-548,686-687,731-828,842-846,864-865,932-969,984-993,996-1018,1033-1043,1046-1064,1095-1145,1167-1189,1387-1471) — Phase 0 GTP-2 confirms zeroget_torch_device()calls in any of them. They are off-limits to this slice.comfy/ldm/seedvr/vae.py,comfy_extras/nodes_seedvr.py,comfy/sd.py— explicitly out of scope per the issue body's Non-Goals.comfy/ldm/seedvr/model.pybut at different sites (structure-resolution helper, RoPE cache); coordinate at merge if both land same day. The two issues edit non-overlapping regions of the same file: select image from list node / or modify SaveImage Comfy-Org/ComfyUI#185 edits MMModule.forward (lines 613–631), Add SeedVR2 support #110 edits the structure-resolution helper and the RoPE cache (different methods per the issue body cross-reference). Neither blocks the other on different methods; the slicer'sgit diff(AC-4) confines hunks to[613, 631]precisely so the merge can be linear.mergeable=DIRTYand silent for two weeks per the issue body; pushing fixes back toyousef-rafat:seedvr2from this issue is out of scope and requires a separate decision.ops,load_device" comment is context, not mandate). The slice removes two lines and adds one test file; it does not propagateload_devicethrough MMModule's constructor or rewire weight placement.MMModule.forwardagainst annn.Identity-shaped stub (_IdentityCallable.forward(x)returnsx); correctness of real SeedVR2 transformer math is covered by the existing seedvr workflow integration onissue_101and is not re-validated by this slice.call_count[0] == 0,torch.equal(...),vid_out.device == vid_in.device, AST[] == []).MMModule.forwardbeyond removing the two cast lines (624–625). The four-linedef forward(...)header through theTuple[...]return type,vid_module = ...,vid = vid_module(vid, ...), theif not self.vid_only:block (including line 629's derived-device cast), and thereturn vid, txtare preserved verbatim.Tracked in pollockjj/mydevelopment Comfy-Org#185 (bookkeeping PR pollockjj/mydevelopment#214).