[OnePunchMonk]

Motivation

Fixes #43107.

SGLDiffusionExecutor._cond_key keyed cached conditioning on the shape plus the first and last scalar of prompt_embeds[0]. On a key hit the executor drops the embeddings and the worker restores whatever is cached under that key. For Flux, prompt_embeds[0] is the pooled CLIP-L y, which ComfyUI takes from the first 77-token chunk only, so two prompts that differ only past that point (or CLIPTextEncodeFlux with the same clip_l text) get the same key and the second prompt runs with the first one's T5 context. Qwen-Image-Edit had the same problem with image_latent, which is dropped but was not part of the key.

Modifications

• _cond_key returns a blake2b digest over prompt_embeds, pooled_embeds, prompt_seq_lens and the adapter's cached extra_req fields. Nested dicts/lists are hashed by content with length framing (no repr() of tensors, which abbreviates large ones).
• Adapters declare cached_extra_keys. drop_cached_fields and the key both read it, so they can't drift apart. For MiniMax-H3 this puts the h3_* fields and sigmas in the key (previously only the text was keyed).
• bind_comfyui_session stores packed extras per comfyui_cond_key instead of in one shared slot. Before, on an H3 cache hit the positive cond could get the negative's h3_context / payload back (last write won). The comfyui_cache_fp check is unchanged.
• Equal content still hits the cache, so later steps keep sending only latents and timestep.
• Tests in test_comfyui_session.py run the real executor bookkeeping plus bind_comfyui_session: Flux same y with different T5, same first/last scalar, different image_latent, H3 pos/neg extras on a cache hit, H3 same text with a different payload, large tensors inside a payload list, per-key extras on the worker, and a repeat that must still use the cache.

Accuracy Tests

No model forward changes. Repro in a real ComfyUI KSampler run (ComfyUI master 58b176f, real CLIP-L weights, plugin FluxExecutor + SGLDModelPatcher wired like core/generator.py, scheduler round-trip replaced by the real bind_comfyui_session), prompt A for 0 to 45%, prompt B from 50%:

before: T5 context the worker used per DiT call: ['A', 'A', 'A', 'A', 'A', 'A']
after:  T5 context the worker used per DiT call: ['A', 'A', 'A', 'B', 'B', 'B']

Unit tests (CPU):

main source + new tests:  7 failed, 11 passed   test_comfyui_session.py
patch: 96 passed, 11 skipped (CUDA only), 3 xfailed   test_comfyui_{session,adapters,h3,generator,profile}.py, test_conditioning_cache.py
patch: 12 passed   apps/ComfyUI_SGLDiffusion/test/test_h3_request.py

Speed Tests and Profiling

The key now copies the embeddings to CPU and hashes them once per step (the old key already synced via .item()). Hashing measured on CPU: about 2.5 ms for a 512x4096 bf16 T5 context, 1.3 ms for a Qwen-Image context. For H3 the payload's cond latents are hashed too. Small next to a DiT step, but I did not measure it on GPU.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations. (not needed)
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed. (CPU only, see above)
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #37803977549
Latest PR Test (Extra): ❌ Run #37803977308
Latest PR Test (AMD ROCm 10): ❌ Run #37803977583

[OnePunchMonk]

Superseded by #43492, which consolidates this fix with 3 related executor conditioning-cache fixes per maintainer request to reduce PR overhead. Closing.