feat(comfyui): switch allocator to cudaMallocAsync (A/B won, fixes LTX OOM)

Mirror comfy-dev's operator-run allocator A/B result off irv-ml1: drop
--disable-cuda-malloc (ComfyUI keeps CUDA's default async allocator) and
remove PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True (native-allocator-only,
inert under cudaMallocAsync). The native+expandable_segments combo was
fragmenting/over-reserving (~45 GB allocated-but-unused) and OOMing the LTX-2.3
v1.5.0 LoRA stack at Gemma TE load; cudaMallocAsync packs tighter + returns
freed blocks so the job fits (stress test peaks ~82% VRAM, 0 OOM). The
shared-A6000 phantom-OOM that --disable-cuda-malloc guarded is gone since TTS
moved to the 3090 (2026-06-18).

stacks/comfyui/compose.yaml

@@ -43,10 +43,6 @@ services:
       - USE_UV=true
       # Extra ComfyUI launch flags (image appends these to main.py, then adds
       # --base-directory + --enable-manager itself):
-      #  --disable-cuda-malloc — use PyTorch's native allocator instead of the
-      #    default cudaMallocAsync, which phantom-OOMs ("allocation would exceed
-      #    allowed memory", 0 B allocated) when the A6000 is shared with the
-      #    pinned TTS services (chatterbox-fast / csm-expressiva, ~22 GB).
       #  --fp8_e4m3fn-text-enc — load the FLUX.2 Qwen3-8B text encoder as fp8
       #    (~8.7 GB) instead of upcasting the fp8 file to fp16 (~16 GB). Matches
       #    the box's Ampere-fp8 posture; the encoder runs once per gen so the
@@ -56,15 +52,23 @@ services:
       #    ComfyUI attention changes"). Binds the in-image sageattention v2.2.0
       #    sm_86 build (rebuilt against the pinned torch 2.12.1). Global speedup
       #    across Flux/SDXL/Wan (comfy-dev benchmarking, 2026-06-18).
-      - COMFY_CMDLINE_EXTRA=--disable-cuda-malloc --fp8_e4m3fn-text-enc --use-sage-attention
-      # Let PyTorch's native caching allocator grow segments instead of
-      # pre-carving fixed blocks — cuts fragmentation OOMs on long-lived
-      # Qwen-Image-Edit sessions (a ~2 GB alloc failing with 1.75 GB free while
-      # 45 GB sat allocated + reserved-but-unallocated). Cache-preserving: packs
-      # better WITHOUT unloading the checkpoint, so no edit-latency hit. Only
-      # valid under the native allocator (paired with --disable-cuda-malloc
-      # above; incompatible with cudaMallocAsync). comfy-dev request 2026-06-25.
-      - PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+      #
+      # ALLOCATOR (2026-07-19 — comfy-dev A/B, operator-run — cudaMallocAsync WON).
+      # We deliberately DO NOT pass --disable-cuda-malloc, so ComfyUI keeps CUDA's
+      # default async allocator (cudaMallocAsync). History: --disable-cuda-malloc +
+      # PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True were added when the A6000 was
+      # SHARED with the TTS zoo (the native allocator dodged a cudaMallocAsync
+      # phantom-OOM). TTS moved to the 3090 (2026-06-18), removing that trigger; the
+      # LTX-2.3 v1.5.0 LoRA stack (DMD+OmniNFT+act LoRAs patching the DiT + the 12B
+      # Gemma text-encoder) then began hitting the 48 GB ceiling under the native
+      # allocator, which fragments/over-reserves (~45 GB allocated+reserved-but-
+      # unallocated) and OOMs at TE load. cudaMallocAsync packs tighter + promptly
+      # returns freed blocks, so the same job now FITS: the operator's previously-
+      # OOMing stress test peaks ~82% VRAM (~40/48 GB) with headroom, 0 OOM/errors.
+      # COUPLING: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True is native-allocator-
+      # only, so it is REMOVED here (inert/invalid under cudaMallocAsync) — revert BOTH
+      # together. If the TTS zoo ever moves back onto the A6000, re-evaluate the pair.
+      - COMFY_CMDLINE_EXTRA=--fp8_e4m3fn-text-enc --use-sage-attention
     volumes:
       - ${COMFYUI_BASEDIR}:/basedir
       # models/ overlaid from storetank. The ~325 GB model tree was migrated