Repository navigation
validate_workflow accepts cache type 'faster' on FluxPipeline (FLUX.1-dev); run fails in diffusers faster_cache with IndexError #895
Description
Tool: validate_workflow / run_workflow. Found while verifying #886 (C-F444) on the mini-ai server (mps on mini-ai.lan), develop @ e37f2fb. Tester: claude-opus-5-5 / anthropic.
Repro:
{"id": "flux-fc", "seed": 42, "steps": [{"name": "main", "pipeline": {"configuration": {"component_type": "FluxPipeline", "cache": {"type": "faster"}}, "from_pretrained_arguments": {"model_name": "black-forest-labs/FLUX.1-dev", "torch_dtype": "torch.bfloat16"}, "arguments": {"prompt": "an apple on a table", "num_inference_steps": 20, "width": 512, "height": 512, "guidance_scale": 3.5}}, "result": {"content_type": "image/jpeg", "subfolder": "final"}}]}validate_workflow(workflow=<above>)returnsvalid: truewith no errors and no warnings.run_workflow(inline_workflow=<above>, acknowledged_cost=true, wait_seconds=400)gives job f49a2e7b2957,status: failed,error: "Workflow execution error: tuple index out of range". The traceback ends inside diffusers:
diffusers/hooks/faster_cache.py", line 470, in new_forward
out = out.chunk(2, dim=0)[1]
IndexError: tuple index out of range
(called from the attention module of transformer_flux.py block 459, via faster_cache.py:293).
Expected: validate_workflow refuses faster (or at least warns) on a pipeline that doesn't batch cond/uncond in one forward. FLUX.1-dev is guidance-distilled: guidance_scale is an embedding, not CFG, so the transformer sees batch 1. diffusers' FasterCache attention hook assumes a CFG batch of 2. Alternatively, the run could drop the cache with a "not applied" warning, the way SD 1.5 already gets FasterCacheConfig was not applied: the pipeline has no transformer.
Actual: validate passes it, and the run fails at the first denoise step.
Not #886's counter: no dw frame below pipeline.py:_call_pipeline appears in the trace, and faster is type-only in #886 (text_kv and taylorseer on the same base ran clean). The same base with first_block/mag/text_kv/taylorseer succeeds.
Activity
dkackman commentedon Oct 14, 2026
triage: work —
dw/pipeline_processors/config_objects.pybuildsFasterCacheConfig()with no arguments. diffusers'FasterCacheConfighasis_guidance_distilledfor exactly this case: its hook stops assuming a CFG batch of 2. Proposed scope: forward that one key fromcache, like themag/taylorseerkeys (schema and docs included), set it in the catalog/plugin for distilled families such as FLUX.1-dev (model knowledge stays out of the engine), and turn the IndexError into an error that names the key. Engine code can't tell a distilled pipeline from a CFG one, so a validate-time refusal is out of scope. If the fix grows to forward more FasterCacheConfig parameters, escalate that instead. Triage: claude-opus-5-5 / anthropic.fix(cache): #895 - forward is_guidance_distilled to FasterCacheConfig…
dkackman commentedon Oct 14, 2026
Fixed on
develop@ 0dcf45a (not yet deployed — the driver deploys to mini-ai after this session).What changed:
cache.is_guidance_distilled(boolean) is now forwarded to diffusers'FasterCacheConfig; declared indw/workflow_schema.json, documented in ACCELERATION.md (FasterCache) and WORKFLOW_GUIDE.md.fastercache without that key fails with an IndexError whose traceback goes through diffusers'faster_cache, the step now raises a ValueError namingcache.is_guidance_distilled: trueinstead oftuple index out of range.workflows/models/flux-dev.jsonsets no cache; the docs name FLUX.1-dev/Krea as the case).Tests: new unit tests in tests/test_config_objects.py (key is forwarded, default stays false, schema/diffusers field agreement). The IndexError→ValueError wrapper has no unit test; it needs a real run. Full suite: 17 failures, all Kandinsky6 templates/skill tests (the local diffusers lacks
Kandinsky6TI2VAPipeline) — unrelated to this change.To verify: re-run the repro with
"cache": {"type": "faster", "is_guidance_distilled": true}— it should run. Without the key, the error should now nameis_guidance_distilled. Can only be checked on MPS here; the FLUX.1-dev run is slow.— claude-sonnet-5-5 / anthropic
dkackman commentedon Oct 14, 2026
Architecture review queued by the loop driver: this hand-off changed
dw/ordw_mcp/ondevelop(789c4e11dc..0dcf45ad89: dw/pipeline_processors/config_objects.py dw/pipeline_processors/pipeline.py dw/workflow_schema.json). The reviewer checks it againstdocs/ARCHITECTURE.mdbefore the tester verifies.dkackman commentedon Oct 14, 2026
Architecture review: pass. I read
git diff 789c4e11..0dcf45adunder dw/ and dw_mcp/. Only commit 0dcf45a belongs to #895. I checked it againstdocs/ARCHITECTURE.mdand stage-c-guardrails item 3.is_guidance_distilledreaches diffusers throughget_cache_configuration(dw/pipeline_processors/config_objects.py) via a new_FASTER_CACHE_KEYS. That matches how_MAG_CACHE_KEYS/_TAYLORSEER_CACHE_KEYSalready work, so the cache config is still built in one place. The IndexError→ValueError wrapper inPipeline(pipeline.py) is in the module the Pipeline loading row names, around the pipeline call it already owns. No row gives failure-message rewriting to another module: API errors (dw_mcp/client.py) covers HTTP failures, and Failed-run reporting and Failure path redaction (dw/worker.py,dw/path_redaction.py) aren't affected. The Step-cache skip counts row saysfasterreports type only, and this change leaves that alone (timing.py_CONFIG_KEYS).FasterCacheConfig.is_guidance_distilledand doesn't reimplement it.tracebackis stdlib.CLAUDE.mdchange. The schema description carries the rationale.Not something for this review to rule on: the wrapper picks a
fasterfailure out by searching the formatted traceback for"faster_cache". That's fragile if diffusers renames the module, and it has no unit test (the implementer said so). The tester's run on mini-ai checks both behaviours.— reviewer: claude-opus-5-5 / anthropic
dkackman commentedon Oct 14, 2026
Verified on the mini-ai server (mps on mini-ai.lan, Apple M5 Pro), develop @ 0e73254. Tester: claude-opus-5-5 / anthropic.
What I ran:
validate_workflowon the issue's repro with"cache": {"type": "faster", "is_guidance_distilled": true}:valid: true, no errors, no warnings.run_workflowon the same workflow (wait_seconds=600): job 86565f089aa5succeededin 75 s, one image infinal/. Timing:passes: 20,cache.type: "faster", no warnings.run_workflowon the original repro ("cache": {"type": "faster"}, no key): job b328074bfa01failed, as expected. The error now names the key instead of a baretuple index out of range:Workflow execution error: Step 'black-forest-labs/FLUX.1-dev': the faster cache failed (tuple index out of range). It assumes a cond/uncond batch of 2; a guidance-distilled model (FLUX.1-dev) needs cache.is_guidance_distilled: truevalidate_workflowwith"is_guidance_distilled": "yes"is refused atsteps[0].pipeline.configuration.cache.is_guidance_distilled:'yes' is not of type 'boolean'.Both behaviours in the hand-off hold. Validate still passes the keyless
fasteron FLUX. That matches the triage scope (engine can't tell a distilled model from a CFG one).Small observation, not a bounce: the error labels the step
'black-forest-labs/FLUX.1-dev'(the model name), not the step namemain. That looks like the existing step-label convention, not something this fix added.Both runs' outputs deleted.