Spectrum MiniMax H3 v0.1.5

Spectrum MiniMax H3 v0.1.5

Fixes MiniMax H3 sampling on Apple MPS and includes the documentation and licensing corrections made since v0.1.4.

Fixed

• Move detached solver timestep tensors to CPU before converting them to float64, avoiding the unsupported MPS float64 conversion that stopped sampling at step 0.
• Apply the same transfer-before-cast ordering to sigma-schedule normalization, which contained the same latent defect.
• Centralize the conversion path so run initialization and per-step coordinate calculation cannot diverge.
• Correct the repository's GPL-3.0-or-later licensing files and package metadata.

Validation

• Add regression coverage that explicitly enforces detach -> CPU transfer -> float64 cast -> flatten.
• Preserve CPU values, output shape, device, and dtype through the shared conversion helper.
• Confirm the fix on a physical Apple M4 Max system with a complete 20-step RES multistep run: 14 actual H3 transformer evaluations, 6 forecasted evaluations, and zero fallbacks.
• Retain the existing native MiniMax H3 fixture, scheduler, transaction, fallback, and forecasting test coverage.

Documentation

• Clarify that Spectrum can produce trajectory deviations and localized degradation during fast or brief motion.
• Record the community-tested Radeon AI PRO R9700 / ROCm configuration and its measured warm-run result as a scoped compatibility datapoint.

Scope

The MPS hardware confirmation covers the reported Apple M4 Max, PyTorch 2.10.0, ComfyUI 0.30.0, RES multistep configuration. It does not establish compatibility for every Apple Silicon model, PyTorch release, sampler, workflow topology, or MiniMax H3 configuration.

Spectrum MiniMax H3 v0.1.4

Spectrum MiniMax H3 v0.1.4

Adds optional device-resident history storage for systems with spare VRAM.

Added

• Add history_storage=system_ram|vram to the Spectrum node, defaulting to the existing system-RAM behavior.
• Keep model-dtype history on the producing device in VRAM mode, avoiding actual-step device-to-host copies and repeated forecast-time host-to-device reads.
• Clone the target view into compact owned device storage so cached history does not retain the complete final-block hidden tensor.
• Report the configured storage and resolved history device in debug summaries.

Highlights

• System-RAM mode remains the default and preserves existing saved-workflow behavior.
• VRAM mode preserves the forecasting math, model dtype, FP32 accumulation order, scheduler, and fallback semantics.
• History is still bounded by max_history and released at run teardown.

Memory guidance

• The supplied 0.5 MP single-branch workflow retained about 2.22-2.27 GiB at max_history=8; a two-branch topology at the same shape would use roughly twice that amount.
• At the native 1344x768, 124-frame example, eight two-branch snapshots approach 6.1 GiB.
• VRAM mode also needs headroom for native model execution, the current snapshot, the prediction result, an FP32 accumulation chunk, and allocator fragmentation.

Measured 0.5 MP A/B results

Three supplied full-checkpoint Euler pairs measured total times of 112.43/105.55 s, 115.60/116.08 s, and 109.80/107.26 s for system RAM/VRAM respectively. The combined means were 112.61 s and 109.63 s, a 2.6% advantage for VRAM mode. Individual pairs ranged from 6.1% faster to 0.4% slower, so the performance benefit is workload- and system-dependent.

Validation

• Automated tests cover setting validation, default compatibility, compact owned storage, storage-device tracking, and CPU/CUDA prediction equivalence.
• Existing native-path equivalence, transaction, sampler-contract, scheduler, and bounded-memory tests remain in place.

Current limits

The supplied 0.5 MP tests verify the expected additional allocation and show a small, variable timing benefit. Other resolutions, durations, CFG topologies, reference modes, and hardware remain unverified. Selecting VRAM storage with insufficient headroom can raise an out-of-memory error.

Spectrum MiniMax H3 v0.1.3

Spectrum MiniMax H3 v0.1.3

Stabilizes deterministic Euler and RES multistep audio-video forecasting and adds Comfy Registry publishing metadata.

Changed

• Require one completed actual H3 evaluation after every RES multistep forecast so its retained old_denoised state is native before forecasting resumes.
• Enforce a three-step actual tail for deterministic RES, including saved workflows that request a smaller tail.
• Require one completed actual H3 evaluation after every Euler forecast to prevent late forecast streaks from accumulating temporal audio/video errors on short schedules.
• Keep ancestral Euler and RES variants on the native path because their injected noise breaks the forecaster's smooth deterministic trajectory assumption.
• Keep tail_actual_steps=1 as the configurable Euler default while applying the sampler-specific RES floor internally.
• Add the xmarre Comfy Registry publisher metadata and publishing workflow.

Highlights

• Forecasts selected post-transformer H3 features while preserving the native current-step output heads, reconstruction, sigma mapping, and return structure.
• Keeps runtime and history state isolated per model clone and rolls incomplete split-branch transactions back to a complete native step.
• Supports deterministic Euler and RES multistep sampling with sampler-specific post-forecast refresh and tail policies, plus explicit native fallback for stochastic samplers, unsupported samplers, incompatible topology, invalid forecasts, and multi-GPU parallel sampling.
• Bounds retained history on CPU and streams forecast accumulation in chunks to avoid persistent full-feature FP32 coefficient or right-hand-side tensors.
• Avoids redundant full-target GPU concatenation and CPU restacking on native single-call actual steps, and reports history/forecast timing counters in debug summaries.
• Leaves the separate FLUX-focused ComfyUI-Spectrum-Proper repository unchanged.

Measured 0.5 MP results

Supplied full-checkpoint, 20-step runs:

Sampler	Native	Spectrum	Schedule	Time saved	Wall-time reduction	Speedup
RES multistep	2:42	1:54	14 actual / 6 forecast	48 s	29.6%	1.42x
Euler	2:38	1:59	13 actual / 7 forecast	39 s	24.7%	1.33x

The tested Euler artifacts appeared fixed. The remaining slight RES artifacts appeared gone after enforcing the three-step actual tail.

Validation

• The local suite passes against native ComfyUI commit e377e263049f9338b4d12a3dd417b36ae62948ff.
• Automated tests cover forecasting mathematics, scheduling, rollback, clone isolation, the actual ComfyUI loader shape, native-path equivalence, and zero transformer-block execution on forecast steps.
• Sampler contract tests verify one model call per deterministic solver step, RES's current/previous-denoised recurrence and update ordering, ancestral noise injection, rollback-safe refresh state, and sampler-specific forecast spacing and tails.

Current limits

The automated environment cannot decode a full MiniMax H3 generation. Broader prompts, durations, resolutions, CFG settings, reference modes, memory peaks, and audiovisual synchronization remain real-generation validation items. Timings can vary with model warmup and GPU state.

Spectrum MiniMax H3 v0.1.2

Spectrum MiniMax H3 v0.1.2

Adds Spectrum forecasting support for ComfyUI's RES multistep sampler family.

Added

• Support sample_res_multistep.
• Support the CFG++, ancestral, and ancestral CFG++ RES multistep variants.
• Cover the sampler allowlist with focused regression tests.

Highlights

• Forecasts selected post-transformer H3 features while preserving the native current-step output heads, reconstruction, sigma mapping, and return structure.
• Keeps runtime and history state isolated per model clone and rolls incomplete split-branch transactions back to a complete native step.
• Supports native Euler, Euler ancestral, and RES multistep sampling, with explicit native fallback for unsupported samplers, incompatible topology, invalid forecasts, and multi-GPU parallel sampling.
• Bounds retained history on CPU and streams forecast accumulation in chunks to avoid persistent full-feature FP32 coefficient or right-hand-side tensors.
• Leaves the separate FLUX-focused ComfyUI-Spectrum-Proper repository unchanged.

Validation

• GitHub Actions passes against native ComfyUI commit e377e263049f9338b4d12a3dd417b36ae62948ff.
• Automated tests cover forecasting mathematics, scheduling, rollback, clone isolation, the actual ComfyUI loader shape, native-path equivalence, and zero transformer-block execution on forecast steps.
• CodeRabbit review feedback was evaluated and all review threads were resolved.

Current limits

Full MiniMax H3 checkpoint generation, reference modes, decoded output quality, audiovisual synchronization, long-duration memory use, VRAM/RSS peaks, and end-to-end speedup remain unverified. Presets and performance expectations are provisional.

Spectrum MiniMax H3 v0.1.1

Spectrum MiniMax H3 v0.1.1

Corrective release for the ComfyUI custom-node entry point.

Fixed

• Load the bundled comfyui_spectrum_h3 implementation through the custom-node package itself, matching ComfyUI's isolated directory loader.
• Preserve direct source-checkout imports used by development and packaging checks.
• Add a regression test that loads the node from an isolated, hyphenated custom-node directory without placing that directory on sys.path.

Highlights

• Forecasts selected post-transformer H3 features while preserving the native current-step output heads, reconstruction, sigma mapping, and return structure.
• Keeps runtime and history state isolated per model clone and rolls incomplete split-branch transactions back to a complete native step.
• Supports native Euler and Euler ancestral sampling, with explicit native fallback for unsupported samplers, incompatible topology, invalid forecasts, and multi-GPU parallel sampling.
• Bounds retained history on CPU and streams forecast accumulation in chunks to avoid persistent full-feature FP32 coefficient or right-hand-side tensors.
• Leaves the separate FLUX-focused ComfyUI-Spectrum-Proper repository unchanged.

Validation

• GitHub Actions passes against native ComfyUI commit e377e263049f9338b4d12a3dd417b36ae62948ff.
• Automated tests cover forecasting mathematics, scheduling, rollback, clone isolation, the actual ComfyUI loader shape, native-path equivalence, and zero transformer-block execution on forecast steps.
• CodeRabbit review feedback was evaluated and all review threads were resolved.

Current limits

Full MiniMax H3 checkpoint generation, reference modes, decoded output quality, audiovisual synchronization, long-duration memory use, VRAM/RSS peaks, and end-to-end speedup remain unverified. Presets and performance expectations are provisional.

Spectrum MiniMax H3 v0.1.0

Spectrum MiniMax H3 v0.1.0

Initial alpha release of the standalone Spectrum-style acceleration node for ComfyUI's native MiniMax H3 audio-video model.

Highlights

• Forecasts selected post-transformer H3 features while preserving the native current-step output heads, reconstruction, sigma mapping, and return structure.
• Keeps runtime and history state isolated per model clone and rolls incomplete split-branch transactions back to a complete native step.
• Supports native Euler and Euler ancestral sampling, with explicit native fallback for unsupported samplers, incompatible topology, invalid forecasts, and multi-GPU parallel sampling.
• Bounds retained history on CPU and streams forecast accumulation in chunks to avoid persistent full-feature FP32 coefficient or right-hand-side tensors.
• Leaves the separate FLUX-focused ComfyUI-Spectrum-Proper repository unchanged.

Validation

• GitHub Actions passes against native ComfyUI commit e377e263049f9338b4d12a3dd417b36ae62948ff.
• 37 automated tests cover forecasting mathematics, scheduling, rollback, clone isolation, native-path equivalence, and zero transformer-block execution on forecast steps.
• CodeRabbit review feedback was evaluated and all review threads were resolved.

Current limits

Full MiniMax H3 checkpoint generation, reference modes, decoded output quality, audiovisual synchronization, long-duration memory use, VRAM/RSS peaks, and end-to-end speedup remain unverified. Presets and performance expectations are provisional.