YuE2-3B on RTX 4070 Ti Super: the 16GB fit is arithmetic, not luck

What You'll Build

A complete song — vocals and accompaniment, 48 kHz stereo — generated on one 16 GB RTX 4070 Ti Super 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 4070 Ti Super (16GB VRAM, Ada Lovelace) · the 16 GB fit is measured on a 16 GB card and every input to it is a property of the model, not the board · See benchmark data

ℹ️ A borrowed measurement is only as good as the reason it transfers, so here is the reason first. The vendor's quick start asks for a 24 GB card and publishes nothing for a 16 GB one, 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 and two that fill the model's context to the token, one of those in the mode that also builds the second guidance branch. No run went out of memory. Those figures are that card's, labelled so wherever they appear. They carry to an RTX 4070 Ti Super not because the two cards are similar but because the quantities involved are not GPU properties at all: the checkpoint size is a byte count, the KV cache is layers × kv-heads × head-dim × dtype × tokens × branches read from config.json, and the ceiling the library imposes is your board's capacity minus 2 GiB. Time is the one thing that does not carry, and none exists for this card. 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. Read those two sentences closely: they constrain a capacity and a dtype, and nothing else — no generation, no bus width, no compute capability. The 24GB is a recommendation rather than a floor the software enforces, which is why somebody opened a request on a low-VRAM inference project the day the model landed, asking "Any possibility to optimise within 16gb vram?" (Wan2GP issue #2283, 2026-09-10). The answer turns out to be that no optimisation is needed, and the following section is the whole argument.

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 no matter what you own, so on a 16 GB board the total - 2 term binds. total is whatever CUDA reports for your board, which is a little under 16 GiB and is not guaranteed to be the same number on two boards of the same nominal size — the measured card reported 15.93 GiB and got a 13.93 GiB ceiling. Read yours from yue2 doctor in step 2 and subtract 2; do not copy that figure. Two consequences hold for every card: 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 contains whatever a display or monitoring tool is holding, and which the clamp cannot see.

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); the same file restates the total as kv_bytes = 2 * n_layers * numel * element_size at L99, which is the arithmetic below in the code's own words. 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. Count the terms: four come from a JSON file, one is the token capacity, one is the branch count. None is a GPU.

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). Both bounds are code, so both are the same on every card that runs this package.

That makes the worst case computable in advance, and it is also the case that was measured:

Against a 16 GiB card that leaves 3.987 GiB for activations, the CUDA context and fragmentation. Those three are the only terms the arithmetic cannot supply, and they are what a measurement is for: on the RTX 5060 Ti 16GB the full-context two-branch run peaked at 12.64 GiB in-process — 0.627 GiB above the floor, with 1.29 GiB still under that card's ceiling. So the derivation predicted the shape and the run filled in the remainder, which is the only part of this page that could have surprised anyone.

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 function hands back a list of such ranges over the frame count, run one after another. 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 name 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 them 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 every line number above is true of either install. Install PyTorch's CUDA build first if your environment would otherwise resolve a CPU-only wheel; beyond that there is no wheel selection to get right on this card.

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

yue2 doctor

Read dependencies_ready and the cuda array in the JSON it prints. That array is where you get the one board-specific number the derivation needs: your total memory in GiB, from which the ceiling is that figure minus 2. It also reports your compute capability, which matters in exactly one place on this page (the FP8 entry in Troubleshooting) and nowhere else. doctor reports environment readiness only, and says so itself — "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, which took 30.40–34.81 s across the four reference runs.

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's framing for throughput is "One song at a time.", and the batch subcommand queues requests rather than running them concurrently.

Results

Measured on a 16 GB card, and that card is not this one

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 — which is why both a device peak and a peak-minus-idle figure appear; the second is what 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. Everything in that sentence is carried by the derivation rather than by the card, which is the reason this page exists at all: were the fit a matter of one board's luck, borrowing it would be indefensible.
• One number in the table is genuinely card-specific, and it is the small one. The gap between the 12.013 GiB floor and the 12.64 GiB peak — 0.627 GiB of activations, CUDA context and fragmentation — is the part no derivation supplies and the part that could differ on an RTX 4070 Ti Super, by tens of megabytes either way. Against 1.29 GiB of remaining headroom on the tighter of the two cards, that is not a margin anyone should worry about; it is, however, the honest answer to "what could be different here".
• Speed: absent for the RTX 4070 Ti Super, deliberately. Seconds belong to the silicon and this is where the two cards genuinely part company. NVIDIA's comparison page lists this card 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 21 Gbps and 28 Gbps respectively (RTX 4070 Ti Super, RTX 5060 Ti). So 672 GB/s against 448 — a factor of 1.50, the narrowest of the GDDR6X cards in this capacity tier and still comfortably in this card's favour. The measured times are therefore an upper bound rather than an estimate. An upper bound is not a measurement. 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 figure 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 are measured against different things, and the run that fills the context to the token fit with 1.29 GiB to spare. The vendor does not define what its maximum-context run was, so 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 two published speed figures, and the reason neither can be scaled to this card

The vendor's reference is an RTX 4090 24GB: 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), method note "NVML records the full-run GPU peak." — the same instrument as the table. The second is the only third-party figure that exists for this runtime: the author of a C++ port measured his engine at a real-time factor of 0.30 against "python 0.31" for the reference implementation on the vendor's example prompt (audio.cpp issue #499, 2026-09-10), naming no card in that comment. Both belong to cards larger than this one, and interpolating between a 24 GB Ada part and a 16 GB Blackwell part to land on a 16 GB Ada part is arithmetic without a mechanism. The memory derivation above has a mechanism; that is the difference, and it is why one of them appears on this page as a conclusion and the other does not.

Nothing is filed at /check/yue2-3b/rtx-4070-ti-super 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 4070 Ti Super clears it. Nothing else in the shipped package tests compute capability on the default path — which is the fact that makes the transfer in this page's opening admonition a claim about code rather than a hope.

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 the mechanism is a nice illustration of a check that asks the wrong question. 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) — that is a question about the operator's registration, not about whether the build compiled its kernel, and a wheel built without USE_FLASH_ATTENTION registers the operator and then throws when it is called. A community ComfyUI wrapper documents exactly that for official Windows wheels: "The op exists and then crashes" (ComfyUI-YuE2 README). The vendor ships the way out: backend="torch-eager" (or --backend torch-eager) is accepted at src/yue2/pipeline.py L128-129 and disables the CUDA-graph path at L250, so the selection never runs. Do not pip install flash-attn; the package declares no such dependency and its "FlashAttention" is PyTorch's own kernel. The derivation above survives the switch — the eager route allocates its per-branch KV cache at len(ids) + max_tokens (src/yue2/sampling.py L76-79), which is at most the graph path's shared capacity, so the KV term cannot exceed the table — but the peak on that path is unmeasured by anyone, as is its speed, which is why this page documents the default. Requesting flash explicitly fails at construction instead of 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 once the device has 2 GiB or less. Run the derivation above on a 12 GB card and you can see why the floor is 16: the 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 not the same claim as a proven failure, and that distinction is exactly why 12 GB stays out rather than being asserted either way.

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 alone, 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

The package has an opt-in FP8 mode for the AR projections, selected with quantization="fp8", and it is the only place in the package that tests compute capability (src/yue2/quantization.py L71-72). NVIDIA's CUDA GPUs list does not enumerate the SUPER SKUs, and every Ada GeForce entry it does list reports 8.9 (CUDA GPUs), so on mechanism the gate should pass here — check the number in yue2 doctor rather than relying on that inference. It changes nothing either way, because the recommendation is to leave the mode off: its 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 derived and 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, and the port landed on that engine's dev branch while this page was being written: re-checked at 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 card, noting that "The numbers are a bit noisy because of background workloads" (vendor issue #163). None of it transfers to the derivation above: 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 says caching more aggressively would take 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 dev 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 SUPER-series card is named, and no other 16 GB board either. Every hardware-shaped issue in the vendor's tracker that a search for "YuE" surfaces — 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.