Repository navigation
validate_workflow passes cache type faster on FLUX.1-dev; run fails needing is_guidance_distilled #897
Description
Tool: validate_workflow vs run_workflow
Server: mps on mini-ai.lan, develop @ 0e73254
Filed by: claude-opus-5-5 via anthropic, while verifying #867 (case C-F444)
Repro
{"id":"flux-fb","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","num_inference_steps":20,"width":512,"height":512,"guidance_scale":3.5}},"result":{"content_type":"image/jpeg","subfolder":"final"}}]}validate_workflow(workflow=<above>)returnsvalid: true, with no errors and no warnings.run_workflow(workflow=<above>, acknowledged_cost=true, workspace="regression-complete")(job eaa66f569193) returnsfailedafter about 37 s:
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: true
Expected
The runtime already knows the rule: FLUX.1-dev is guidance-distilled, and faster needs is_guidance_distilled: true. validate_workflow should report it, as an error or at least a warning, so the user isn't charged a model load to find out. Adding "is_guidance_distilled": true makes the run succeed (job 05804e87211d).
Note
A related gap: cache: {"type":"first_block"} validates clean on SD1.5 (UNet) and on Z-Image, then at run time only warns that it was "not applied" (jobs c91c3bee99cd, fb19b96cd5a7). It's worth considering in the same place: validate knows the component type.
Activity
dkackman commentedon Oct 14, 2026
triage: escalate — this asks for a new
validate_workflowrule matrix (cache type × pipeline/model:fasteron a guidance-distilled model,first_blockon UNet and on transformers diffusers doesn't hook). #895's triage left the FLUX half out of scope because the engine can't tell a distilled model from a CFG one without putting model knowledge in engine code. That half needs a decision on where the rule lives (catalog/plugin vs engine). The first_block half is a new per-component-class check. Both need sign-off before a fix session builds them. Triage: claude-opus-5-5 / anthropic.dkackman commentedon Oct 14, 2026
Approved (Don, 2026-10-14): option 1. No validate-time rule matrix.
Scope:
enable_cache_on_transformer(dw/pipeline_processors/components.py), next to thecurrent_timestep_callbackwiring: when the config is aFasterCacheConfigand the workflow didn't setis_guidance_distilled, set it from the loaded transformer'sconfig.guidance_embeds. Log at info when it is inferred. An explicit workflow value always wins.guidance_embedsmisses one, say which in the hand-off; don't add model names to the engine.The first_block half is split out to #904.
dkackman commentedon Oct 14, 2026
triage: work — Don's approved scope (option 1) is the whole job: in
enable_cache_on_transformer(dw/pipeline_processors/components.py), inferis_guidance_distilledfrom the transformer'sconfig.guidance_embedswhen a FasterCacheConfig doesn't set it. Keep #895's message (0dcf45a) as the fallback. It isn't on develop yet. #904 is a separate validate/registry change and doesn't share code with this. Triage: claude-opus-5-5 / anthropic.fix(cache): #897 - infer FasterCache is_guidance_distilled from trans…
dkackman commentedon Oct 14, 2026
Fixed on develop @ 5600482 (merge of fix/897-faster-cache-distilled). Not yet deployed: the driver deploys develop to mini-ai after this session.
What changed:
enable_cache_on_transformer(dw/pipeline_processors/components.py) now, for aFasterCacheConfigwhose workflow did not setis_guidance_distilled, sets it True when the loaded transformer'sconfig.guidance_embedsis True, and logs at info when it infers. An explicit workflow value (true or false) always wins; the caller passes whether the key was present. #895's IndexError→ValueError message remains the fallback. No model names in the engine. Unit tests intests/test_pipeline_components.pycover infer / no-embeds / explicit-wins; ruff and the cache tests pass.To verify: the keyless repro in the issue (FLUX.1-dev,
cache: {type: faster}, nois_guidance_distilled) should now succeed. Server log should show "Inferring FasterCache is_guidance_distilled=True".FLUX.1-schnell: its transformer has
guidance_embeds: false, so nothing is inferred there. I haven't run it. By code reading, FluxPipeline only batches CFG when a negative prompt/true_cfg_scale>1 is given, so schnell likely hits the same batch-of-2 assumption; if the tester tries it and it fails, that case isn't covered byguidance_embedsand the user must set the flag. (Not tested here.)claude-sonnet-5-5 via anthropic.
dkackman commentedon Oct 14, 2026
Architecture review queued by the loop driver: this hand-off changed
dw/ordw_mcp/ondevelop(c243b9945b..560048269d: dw/pipeline_processors/components.py dw/pipeline_processors/pipeline.py). The reviewer checks it againstdocs/ARCHITECTURE.mdbefore the tester verifies.dkackman commentedon Oct 14, 2026
Architecture review: pass. Read
git diff c243b994..56004826underdw/anddw_mcp/, limited to #897's commit c23b610 (dw/pipeline_processors/components.py,dw/pipeline_processors/pipeline.py). Checked it againstdocs/ARCHITECTURE.mdand item 3 ofstage-c-guardrails.md.enable_cache_on_transformer, next to the existing FasterCachecurrent_timestep_callbackwiring.pipeline.pyonly passes whether the workflow set the key. Validation isn't touched, so this doesn't overlap the "Acceleration warnings" row (dw/acceleration_warnings.py).transformer.config.guidance_embedsand sets a field on diffusers'FasterCacheConfig.CLAUDE.mdchange.No
arch-approvedlabel, so nothing was waived. Over to the tester for the mini-ai repro.Reviewer: claude-opus-5-5 via anthropic.
dkackman commentedon Oct 14, 2026
Verified on mps (mini-ai.lan), develop @ 29d4c4c. Tester: claude-opus-5-5 via anthropic.
cache: {"type":"faster"}, nois_guidance_distilled, workspaceregression-complete):validate_workflowis valid, andrun_workflow(job 65b710355470) now succeeds. One image came out; manifesttiming.cache.type: "faster", 20 passes, denoise 23.1 s."is_guidance_distilled": false(job b4ffd1983493) fails with validate_workflow accepts cache type 'faster' on FluxPipeline (FLUX.1-dev); run fails in diffusers faster_cache with IndexError #895's fallback message, unchanged: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: true. Nothing was inferred over the explicitfalse.Not observable over MCP: the info line "Inferring FasterCache is_guidance_distilled=True" doesn't show up in
get_job_events(kinds=["log","warning"])for job 65b710355470. The events hold only workflow and file log lines, so I can't confirm it from here. That's not a blocker. The behavior it describes is confirmed.Not tried: FLUX.1-schnell (the hand-off's open question). Both outputs deleted.