YuE2-3B on RTX 4080: the measured 16GB fit, carried across to Ada

What You'll Build

A complete song — vocals and accompaniment, 48 kHz stereo — generated on one 16 GB RTX 4080 from a style prompt and lyrics, together with the editable ABC score the model planned it from, so you can revise the melody or harmony and re-render the same song.

Hardware data: RTX 4080 (16GB VRAM, Ada Lovelace, compute capability 8.9) · the 16 GB fit is measured, on a card of a different generation and the same capacity · See benchmark data

ℹ️ This is a cross-generation clone of a measurement, and the seam is worth seeing. The vendor's quick start asks for a 24 GB card and publishes nothing for any 16 GB board, so on 2026-09-10 the site operator ran YuE2-3B four times on an RTX 5060 Ti 16GB — two songs at the shipped semantic budget, two more filling the model's context to the token, one of those in the mode that also builds the second guidance branch. Nothing ran out of memory. That card is Blackwell; this one is Ada, a generation older. What crosses the seam is the memory result, and you can check why rather than take it: the two tensors that decide the peak are sized from the model's config.json and from the library's own process clamp, and the clamp reads the board's capacity, never its architecture. What does not cross is time, and this card is faster on the axis that matters — so the measured seconds are an upper bound here and not an estimate. If you run this pair, send us your numbers.

⚠️ The weights are non-commercial. YuE2's code is Apache 2.0, but the checkpoints this recipe downloads are CC BY-NC 4.0. The repository's LICENSE scopes that to "the YuE2 checkpoint weights in model.safetensors, or the corresponding" model files. The licence text names the weights and does not address the audio you generate with them — if you intend to release or monetise output, read the full licence rather than this paragraph.

ℹ️ One runtime, and it is a Python wheel. YuE2-3B runs through the vendor's own yue2 package with CUDA. There is no llama.cpp, Ollama or LM Studio path; a GGUF conversion exists for a different engine and its status changed on the day this page was written — see Troubleshooting.

Requirements

The vendor states its hardware line as "Linux · Python 3.10+ · 24GB NVIDIA GPU with BF16 support." on the model card, and as "NVIDIA GPU with BF16 support and 24 GB VRAM." in the GitHub repository. Both name a capacity; neither names a generation, an architecture or a compute capability. That 24GB is a recommendation rather than a floor the software enforces, and it is why someone opened a request on a low-VRAM inference project the day the model appeared, asking "Any possibility to optimise within 16gb vram?" (Wan2GP issue #2283, 2026-09-10). No optimisation is required. The arithmetic follows, then the runs that confirm it.

Why 16 GB holds, in bytes — and why a longer song cannot change it

The library caps the process, not the card. YuE2Pipeline.__init__ reads the device's total memory and sets budget = min((memory_budget_gib - 2) * 2**30, total - 2 * 2**30), then enforces it with set_per_process_memory_fraction (src/yue2/pipeline.py L161-165). memory_budget_gib defaults to 24 whatever you are running on, so on a 16 GB board the total - 2 term binds — and note what that expression contains: a capacity, and nothing else. No branch of it asks which generation the silicon belongs to. total is what CUDA reports for your particular board, a little under 16 GiB and not necessarily identical across boards of the same nominal size; the measured card reported 15.93 GiB for a ceiling of 13.93 GiB. Read yours from yue2 doctor in step 2 and subtract 2. Two consequences follow everywhere: an out-of-memory error can arrive with memory physically free on the board, and the figure to compare against the ceiling is the process's own peak — not the device peak, which also holds whatever a display or monitoring tool is using and which the clamp never sees.

The KV cache is sized by the budget you ask for, not by how many tokens the model emits. The graph runner preallocates, per layer and once each for keys and values, a tensor shaped (branches, capacity, num_key_value_heads, head_dim) where capacity = max(len(prefix)) + max_tokens (src/yue2/cuda_graph.py L70, L90-L92). From config.json — 28 layers, 8 key/value heads, head dimension 128, dtype bfloat16 — one token on one branch costs 28 × 8 × 128 × 2 bytes × 2 tensors = 114,688 B, i.e. 112 KiB. At the full context that is 2.625 GiB on one branch and 5.25 GiB on two. Every factor there is read out of a JSON file that ships with the weights.

Two is the hard maximum, and so is the context. The sampler builds either [prefix] or [prefix, negative] — one branch at guidance exactly 1, two otherwise, and never more (src/yue2/sampling.py L92); cot="off" takes the two-branch path because its default guidance is 1.01 (src/yue2/protocol.py L106). A request whose prefix plus budget exceeds 24,576 tokens is rejected outright rather than quietly trimmed — "Prefix + requested generation budget exceeds 24576; no implicit truncation" (src/yue2/sampling.py L62-63) — and the context is not tunable: the generation config raises "Require context=24576 and midpoint with positive integer steps" for any other value (src/yue2/protocol.py L54-55).

So the worst case is bounded, and the bound is the case that was measured:

Against a 16 GiB card that leaves 3.987 GiB for activations, the CUDA context and fragmentation — and the measurement landed where that predicts: on the RTX 5060 Ti 16GB the full-context two-branch run peaked at 12.64 GiB in-process, 0.627 GiB above the arithmetic floor, 1.29 GiB short of that card's ceiling.

A longer song cannot exceed it. Song length past the context limit is realised as more NAR and VAE chunks, not bigger ones: each chunk's width is computed from the context and the prefix, size = min((context - prefix_tokens - 3) // 2, CONTEXT) (src/yue2/protocol.py L141-145), and the chunks are processed sequentially. More audio therefore costs more time, not more memory, and the two tensors that dominate the peak are already at their maximum above.

Installation

1. Install the inference package into a dedicated virtual environment

The venv is not a style preference. The wheel pins all eight of its runtime dependencies with ==, torch==2.10.0 among them, and an == pin does not step aside for a newer version that is already installed — it replaces it. A ComfyUI user reported on 2026-09-10 that installing this wheel into a working portable build rewrote PyTorch 2.11+cu130 down to 2.10 and left the launcher failing to start. There is also no yue2, yue2-infer or yue2_infer package on PyPI, and the vendor's own setup notes say what not to install: "unverified package with a similar name from PyPI." (skills/yue2-music/references/models-and-setup.md).

python3 -m venv .yue2 && source .yue2/bin/activate
python -m pip install -U pip huggingface-hub==0.36.2
hf download m-a-p/YuE2-3B yue2_infer-0.1.5-py3-none-any.whl --local-dir .
python -m pip install ./yue2_infer-0.1.5-py3-none-any.whl

Those middle three lines are the model card's quick start unchanged, and 0.1.5 is the version the card names — the same wheel the runs below were taken on (sha256 8801e2c0…, 66,117 bytes). The repository tree has moved on to 0.1.6; unpacking both and comparing file by file leaves 12 of 14 modules byte-identical, the differences confined to the version string and to cli.py. Every engine module cited on this page — pipeline.py, cuda_graph.py, sampling.py, protocol.py, storage.py, quantization.py — is identical in both, so nothing here depends on which one you install. Install PyTorch's CUDA build first if your environment would otherwise resolve a CPU-only wheel; there is no special wheel selection to make beyond that.

2. Confirm the card, the dependencies, and your own ceiling

yue2 doctor

Read dependencies_ready and the cuda array in the JSON it prints. The array carries your board's name, its total memory in GiB and its compute capability — subtract 2 from that total and you have the ceiling the clamp will hold this process to. An RTX 4080 reports 8.9, which is the figure NVIDIA lists for it (CUDA GPUs); the measured card reports 12.0, and the only place in the whole package where that difference matters is an opt-in mode covered in Troubleshooting. doctor reports environment readiness only, and is explicit about it in its own output — "Environment readiness is not quality or real-24GB acceptance." (src/yue2/cli.py L78).

3. Let the loader fetch the weights

The first pipeline call downloads what it needs from m-a-p/YuE2-3B and m-a-p/YuE2-Vae. The loader uses an explicit allow-list (src/yue2/storage.py L32-40) rather than a full clone, so the demo audio and images in those repositories are skipped — that is why the Storage row is 7.79 GB and not the 7.83 GB the two repositories hold in total. Weight files are hash-checked against weights_manifest.json as they load. Loading took 30.40–34.81 s across the four reference runs, hash verification included.

Running

Load the pipeline once, then generate:

import json
from pathlib import Path
from huggingface_hub import hf_hub_download
from yue2 import YuE2Pipeline

repo = "m-a-p/YuE2-3B"
demo = json.loads(Path(hf_hub_download(repo, "examples/tonight-awake.json")).read_text(encoding="utf-8"))

pipe = YuE2Pipeline.from_pretrained(repo, device="cuda")
song = pipe(style=demo["style"], lyrics=demo["lyrics"], cot="full", seed=demo["seed"])
song.save("song.flac")
song.save_artifacts("outputs/song")  # ABC, tokens, latents, audio and settings

That is the exact call the two shipped-budget runs below were measured on, with the vendor's own example prompt — which ships inside the model repository and is fetched separately from the weights, because the loader's allow-list does not include it. The two context-filled runs drove the same pipeline one stage at a time (plan → generate_semantic → synthesize → decode) purely so the semantic budget could be set to 24576 - len(plan.prefix) exactly; every other sampling parameter stayed at its shipped default.

song.flac is the finished stereo song. outputs/song holds the ABC score, the semantic tokens, the acoustic latents and the settings used — edit score.abc and pass it back as abc= to re-render the same song with a revised melody or harmony.

The same modes are available from the shell, which is the easier path when you want to queue several songs:

yue2 generate --cot full --style "Mandarin funk, nu-disco" --lyrics-file lyrics.txt --output outputs/song

cot="full" plans melody and chords and is the default; cot="melody" plans melody only and is what the vendor recommends for covers; cot="off" generates with no symbolic plan and, as derived above, is the mode that costs a second KV branch. Do not pass --budget on a 16 GB card: the board already clamps the default, so 16 changes nothing, and 12 or below lowers the ceiling to 10 GiB while halving the VAE decode tile (src/yue2/pipeline.py L148). The vendor frames this as a serial workload — "One song at a time." — and the batch subcommand queues requests rather than running them concurrently.

Results

Measured on a 16 GB card, of a different generation

All four rows were run by the site operator on 2026-09-10 on an RTX 5060 Ti 16GB (Blackwell, compute capability 12.0, driver 591.86, WSL2 Ubuntu 24.04, yue2_infer 0.1.5). Instrumentation was whole-device NVML sampled at 20 Hz — the same instrument the vendor used, and deliberately not torch.cuda.max_memory_allocated(), which excludes both the CUDA context and the allocator's reserved-but-unused blocks and therefore reads low. Each run was a separate process, because PyTorch's caching allocator returns nothing to the driver on its own: after del and gc.collect() the pipeline still held 7.92 GiB, and only torch.cuda.empty_cache() brought it to 1.14 GiB. That card was not headless — a display held 0.76–0.84 GiB throughout — hence both a device peak and a peak-minus-idle figure, the latter being the one the clamp applies to. One rig, one operator, one run per configuration.

• VRAM usage: 9.03 GiB in-process for an ordinary song and 12.64 GiB in the worst case the library's own allocation rule permits — full context on both guidance branches. The closest any run came to its ceiling was 1.29 GiB, and none of the four failed. That is the half of the result which survives the change of generation, because none of its inputs is architectural: the checkpoint and the KV cache are sized from config.json and the requested token count, and the ceiling is total - 2 GiB on any board. An RTX 4080's own peak has never been observed, and any board is entitled to differ by the tens of megabytes that the CUDA context and fragmentation account for.
• Speed: no figure is given for the RTX 4080, and the reason is not caution for its own sake. Seconds are a property of the silicon, and here two things change at once. NVIDIA's comparison page lists the RTX 4080 as 16 GB GDDR6X on a 256-bit interface against the measured card's 16 GB GDDR7 on 128-bit (compare specs), and board-partner tech specs put the pins at 22.4 Gbps and 28 Gbps respectively (RTX 4080, RTX 5060 Ti). So the memory generation moves one step back while the bus width doubles, and the doubling wins outright: 716.8 GB/s against 448 GB/s, a factor of 1.60. For an autoregressive decode that is the axis that dominates, so the measured times are an upper bound for an RTX 4080 — but an upper bound across a generation change is a weaker statement than one within a generation, and it is not a measurement of anything. Send us the real number.
• The context-filled rows are the point of this page. The vendor reports that "maximum-context testing peaked at" 14.08 GiB on its own hardware — a number which, read as a device peak against a 16 GB card's process ceiling, appears to rule out a maximum-length song. It does not: the two figures are measured against different things, and the run that fills the context to the token fit with 1.29 GiB to spare. Because the vendor never defines what its maximum-context run was, this page claims only to have measured the worst case the shipped code can be asked for.
• Quality notes: the vendor reports a 6.7316 SongBench average for YuE2 and 6.9632 for its best-of-8 setting across 192 WildSongBench prompts, using the legacy decoder — its own automatic metrics under its own candidate-selection protocol, not an independent evaluation. Pass vae="m-a-p/YuE2-Vae-legacy" to from_pretrained to reproduce that protocol; the default decoder is m-a-p/YuE2-Vae, which is what the rows above used.

The vendor's own Ada card is the nearest published reference, and it has 24 GB

The vendor benchmarked an RTX 4090 24GB — the same Ada generation as this card, with more of everything — publishing 139.48 LM tokens/s, 214.85 s of audio in 71.04 s and an 11.18 GiB NVML peak in cot="full" mode, headlined as "A 3.6-minute song in 71 seconds on an RTX 4090." (model card). Its method note reads "NVML records the full-run GPU peak.", the same instrument as the table above. Do not read that row as an approximation of this card: it is a larger die on a wider bus with a third more memory, and the vendor names no other Ada part. The only other published figure for this runtime anywhere is a bare real-time factor — the author of a C++ port measured his own engine at 0.30 against "python 0.31" for the reference implementation on the vendor's example prompt (audio.cpp issue #499, 2026-09-10) — with no card named in that comment, no mode, no methodology and no memory figure attributable to the Python side.

Nothing is filed at /check/yue2-3b/rtx-4080 yet, so that page is where the first measurement of this pair will land — contribute yours.

Troubleshooting

The unquantized preset requires CUDA BF16 support

This is the pipeline's only hard hardware gate, and it is a BF16 test rather than an architecture test: the constructor calls torch.cuda.is_bf16_supported() and raises that message when it returns false (src/yue2/pipeline.py L158-160). Ada supports BF16 and an RTX 4080 clears it. Nothing else in the shipped package tests compute capability on the default path, which is the single most important sentence on this page for anyone cloning a Blackwell measurement onto an older card.

An out-of-memory error with memory free on the card

That is the clamp, not the board. set_per_process_memory_fraction holds the process to total - 2 GiB while nvidia-smi still shows free memory — the reference card's two-branch context-filled run peaked with 2.45 GiB physically free and only 1.29 GiB under its own 13.93 GiB ceiling. Two things make it worse and both look like helping: passing --budget 12 (or lower) caps the process at 10 GiB and halves the VAE decode tile, and running a second generation inside the same Python process leaves the previous run's allocations resident, because PyTorch's caching allocator does not return them to the driver. On the reference machine a loaded pipeline held 7.92 GiB, still 7.92 GiB after del and gc.collect(), and 1.14 GiB after torch.cuda.empty_cache(). Call torch.cuda.empty_cache() between songs, or use one process per song.

RuntimeError: USE_FLASH_ATTENTION was not enabled for build. on native Windows

The reference runs were made under WSL2 on Linux wheels; a native Windows install hits this instead, and nothing about it is Ada-specific. The graph runner's flash-attention eligibility test asks whether torch.ops.aten._flash_attention_forward exists and whether its schema carries seqused_k (src/yue2/cuda_graph.py L75-76) — a question about the operator's registration, not about whether the build compiled its kernel. On a wheel built without USE_FLASH_ATTENTION the operator is registered and then throws when called, which is exactly what a community ComfyUI wrapper documents for official Windows wheels: "The op exists and then crashes" (ComfyUI-YuE2 README). The vendor's own escape hatch is a documented backend, not a patch: backend="torch-eager" (or --backend torch-eager) is accepted at src/yue2/pipeline.py L128-129 and switches off the CUDA-graph path at L250, so the selection above never runs. Do not pip install flash-attn to fix it: the package declares no such dependency, and its "FlashAttention" is PyTorch's own kernel. Two caveats are why this page still documents the default path — the eager route allocates its per-branch KV cache at len(ids) + max_tokens (src/yue2/sampling.py L76-79), so the bound derived above still holds and cannot be exceeded, but its peak is unmeasured by anyone and so is its speed. Asking for flash explicitly fails at construction rather than at the first token: attention_backend="flash" raises "Pinned PyTorch variable-length CUDA FlashAttention is unavailable" (src/yue2/cuda_graph.py L82-83).

Memory budget must leave room for a 2GiB reserve

The same clamp on a card too small for it: budget comes out at zero or below when the device has 2 GiB or less. It also explains why the floor here is 16 GB and not lower. A 12 GB card's ceiling works out to 10 GiB — both terms of the min evaluate there, so no --budget value raises it — against the vendor's own 11.18 GiB ordinary-case peak, with the context not reducible and the two candidate levers (--budget 12 --offload-ar) unmeasured by anyone. A card with no documented headroom is a different claim from a proven failure, and that is precisely why 12 GB stays out until somebody runs it.

Raising cfg_scale costs a second KV branch

The model card's tuning table suggests cfg_scale=1.2 to strengthen text guidance, and any guidance other than 1 makes the sampler build a second, unconditional branch whose count is the leading dimension of the KV cache — 2.625 GiB more at full context, as derived above. On 16 GB it is affordable, and the two-branch case is measured rather than assumed: both cot="off" rows run two branches, and the context-filled one is the largest KV allocation the code permits, so raising cfg_scale cannot push that term past the table. What those rows cannot price is the second branch on its own, because cot="off" also skips the ABC planning stage entirely. Change one thing at a time and watch the peak.

Experimental FP8 AR requires CUDA compute capability >=8.9

Here is the one place where Ada and Blackwell are not interchangeable, and it lands closer than it looks. The package has an opt-in FP8 mode for the AR projections, selected with quantization="fp8", and it is the only compute-capability test in the whole package: if torch.cuda.get_device_capability(device) < (8, 9): raise (src/yue2/quantization.py L71-72). An RTX 4080 reports exactly 8.9, so it passes the gate with nothing to spare, where the measured Blackwell card clears it by three whole versions. Leave the mode off regardless: the module's own status output calls its quality unvalidated, it is incompatible with the CUDA-graph decode path, and it keeps a second copy of the BF16 weights in host memory. The unquantized model fits, as measured.

The GGUF files on the Hub, and the C++ engine that reads them

A GGUF conversion of this model exists at audio-cpp/Yue2-3B-GGUF, and it is not a llama.cpp or Ollama artifact — its header declares general.architecture = "audiocpp", which is not among llama.cpp's registered architectures. It targets 0xShug0/audio.cpp, a separate ggml-based audio engine. That port landed on the engine's dev branch while this page was being written, and was still there on re-check: head 3caeba87, committed 2026-09-10 at 17:09 UTC, 29 yue2 paths including src/models/yue2/ and docs/models/yue2.md. Its author reports a peak below 9 GB and a real-time factor of 0.23 to 0.28 through that engine's interface, on a 32 GB Blackwell card, and says of his own figures that "The numbers are a bit noisy because of background workloads" (vendor issue #163). Two qualifications before that sub-9 GB peak changes anything for you: it is a different engine running a Q8 quantisation, self-reported by the person who wrote the port and unverified by us, and it is the deliberately smaller of his configurations — in the same thread he notes that caching more aggressively would push the real-time factor below 0.2 at higher VRAM. By his own announcement the model sits in a branch for community testing that "won’t be included in the prebuilt binaries before merging" (discussion 3). Build the dev branch if you want it, and expect to be an early tester. This page documents the vendor's Python runtime, which is what every figure above was measured on.

Community reports so far

YuE2-3B was published on 2026-09-09. Re-checked on 2026-09-10 at 21:00 UTC: m-a-p/YuE2-3B carries three discussion threads, all opened the previous day — a question about getting reliably instrumental output, a ComfyUI node-pack announcement, and the audio.cpp port above — and m-a-p/YuE2-Vae carries none. The same query against the org's YuE v1 model returns ten threads, so the near-silence is the model's age rather than a broken check. A GitHub-wide issue search for YuE2 returns 22 items: the vendor's own release pull requests, three requests to port the model to other engines, and two bugs in a community ComfyUI wrapper. Reading the title and body of all 22, all three discussion threads in full, and the comment threads on the port requests: no Ada card other than the vendor's own RTX 4090 is named, and no 16 GB card of any generation appears at all. Every hardware-shaped issue you will find in that tracker by searching for "YuE" — a ROCm report, Apple Silicon requests, GGUF requests — predates YuE2 and was filed against YuE v1, a different model with a different runtime; do not transfer their answers here. Nothing in the current package mentions ROCm, HIP or Metal, and the vendor documents no AMD path.

If you hit something, please report it via the submission form.