YuE2-3B on RTX 5070 Ti: same pins as the measured card, twice the bus

What You'll Build

A finished song — vocals and accompaniment, 48 kHz stereo — produced on a single 16 GB RTX 5070 Ti from a style prompt and a lyric sheet, alongside the editable ABC score the model planned it from, so a revised melody or harmony can be re-rendered as the same song.

Hardware data: RTX 5070 Ti (16GB VRAM, Blackwell, compute capability 12.0) · the 16 GB fit is measured, on a card of the same architecture running the same 28 Gbps memory on half the bus · See benchmark data

ℹ️ What is measured here, what is carried, and what is simply absent. 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, and two more that fill the model's context to the token — one of those in the mode that also builds the second guidance branch, which is the largest allocation the shipped code can be asked for. No run went out of memory. Every figure in Results is that card's and is labelled so. What carries to an RTX 5070 Ti is the memory verdict, and for a checkable reason: the two tensors that decide it are sized from the model's config.json and from the library's own process clamp, neither of which reads the GPU, and both cards hold 16 GB. What is absent is time. This card and the measured one are the same silicon generation clocked on the same 28 Gbps memory chips, but this one wires twice as many of them, so its seconds differ and nobody has recorded them. 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 and neither names an architecture — and that 24GB is a recommendation, not a floor the software enforces. It is also the reason somebody opened a request on a low-VRAM inference project within hours of the release, asking "Any possibility to optimise within 16gb vram?" (Wan2GP issue #2283, 2026-09-10). Nothing needs optimising. Here is the arithmetic, and then the runs that bear it out.

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 on every card, so on a 16 GB board the total - 2 term is the one that binds. total is whatever CUDA reports for your board — a little under 16 GiB, and not identical across boards of the same nominal size. The measured card reported 15.93 GiB, which put its ceiling at 13.93 GiB; read your own out of yue2 doctor in step 2 and subtract 2 rather than assuming that number is yours. Two consequences hold 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 carries whatever a display or monitoring tool holds 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 in that product is the model's; none of them is the board's.

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 itself 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), and config.json agrees with a max_position_embeddings of 24,576.

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

Against a 16 GiB card that leaves 3.987 GiB for activations, the CUDA context and fragmentation — and the measurement lands 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, still 1.29 GiB under 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 comes from the context and the prefix, size = min((context - prefix_tokens - 3) // 2, CONTEXT) (src/yue2/protocol.py L141-145), and the chunks run one after another. So more audio 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 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 the hazard directly: "unverified package with a similar name from PyPI." is the thing not to install (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 since moved 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 nothing here depends on which wheel you install. Install PyTorch's CUDA build first if your environment would otherwise resolve a CPU-only wheel; on the reference machine the pin resolved to 2.10.0+cu128, and cu128 is the build that carries the sm_120 kernels a 50-series card needs.

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 gives your board's name, its total memory in GiB and its compute capability; subtract 2 from the total and you have the ceiling this process will be held to. NVIDIA lists the whole 50-series at compute capability 12.0 (CUDA GPUs), so an RTX 5070 Ti should report 12.0 here. Note what doctor does not claim, in its own words — "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 — which is why the Storage row reads 7.79 GB rather than the 7.83 GB the two repositories hold between them. Weight files are hash-checked against weights_manifest.json as they load. That took 30.40–34.81 s across the four reference runs, hash verification included; expect less on a card with more memory bandwidth, by an amount this page cannot quantify.

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, using 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 stage by stage (plan → generate_semantic → synthesize → decode) for one reason only: so the semantic budget could be set to 24576 - len(plan.prefix) exactly. Every other sampling parameter was left 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 easier 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, per the section above, is the mode that costs a second KV branch. Leave --budget alone on a 16 GB card: the board already clamps the default, so 16 changes nothing, while 12 or below lowers the ceiling to 10 GiB and halves the VAE decode tile (src/yue2/pipeline.py L148). Treat this as a one-song-at-a-time workload — the vendor's own framing 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 it was 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). The instrument was whole-device NVML sampled at 20 Hz — the same one 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 so reads low. Each run was its own process, because PyTorch's caching allocator hands nothing back to the driver by itself: after del and gc.collect() the pipeline still held 7.92 GiB, and only torch.cuda.empty_cache() brought it down to 1.14 GiB. A display was attached to that card and held 0.76–0.84 GiB throughout, which is why both a device peak and a peak-minus-idle figure appear below — the second is 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 allocation rule permits — full context on both guidance branches. The closest approach to a ceiling was 1.29 GiB, and no run failed. This is the half that carries to an RTX 5070 Ti: 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, so a 16 GB card of any generation lands on the same arithmetic. This card'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: deliberately absent for the RTX 5070 Ti, and here is the whole argument. Seconds belong to the silicon, and the difference between these two cards is unusually easy to state. NVIDIA's comparison page lists both as 16 GB GDDR7, the RTX 5070 Ti on a 256-bit interface and the measured card on 128-bit (compare specs); board-partner tech specs put both at 28 Gbps per pin (RTX 5070 Ti, RTX 5060 Ti). So 896 GB/s against 448 — exactly twice — with the same architecture, the same compute capability and the same memory chips on both sides. Nothing in that comparison is ambiguous about direction, which makes the measured times a ceiling for this card rather than an estimate of it. A ceiling is still not a measurement; send us the real one.
• 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, looks like it rules out a maximum-length song. It does not: the two numbers are measured against different things, and the run that filled the context to the token cleared it with 1.29 GiB 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, not to have reproduced theirs.
• 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 what neither of them is

One is the vendor's. Its reference card is an RTX 4090 24GB, where it publishes 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), with a method note reading "NVML records the full-run GPU peak." — the same instrument as above. The other is the only third-party figure for this runtime that exists: the author of a C++ port, comparing his engine against the reference implementation on the vendor's own example prompt, put his at a real-time factor of 0.30 and the Python runtime at "python 0.31" (audio.cpp issue #499, 2026-09-10). That comment names no card, no mode and no method, and the memory figure beside it is not attributable to the Python side; his other numbers in that thread are on an RTX 5090. Both bracket what a large card does. Neither says anything about a 16 GB one.

Nothing is filed at /check/yue2-3b/rtx-5070-ti yet, so that page is where the first measurement of this pair will land — contribute yours.

Troubleshooting

The unquantized preset requires CUDA BF16 support

The pipeline's only hard hardware gate, and it tests BF16 rather than an architecture: the constructor calls torch.cuda.is_bf16_supported() and raises that message when it returns false (src/yue2/pipeline.py L158-160). An RTX 5070 Ti passes it, as does any NVIDIA generation with BF16 support. Nothing else in the shipped package tests compute capability on the default path.

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

That is the clamp rather than the board. set_per_process_memory_fraction holds the process to total - 2 GiB while nvidia-smi still reports free memory: on the reference card the 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 moves make it worse and both look helpful. Passing --budget 12 or lower caps the process at 10 GiB and halves the VAE decode tile. Running a second generation inside the same Python process leaves the first run's allocations resident, because the caching allocator does not return them to the driver — measured on the reference machine as 7.92 GiB held by a loaded pipeline, still 7.92 GiB after del and gc.collect(), and 1.14 GiB after torch.cuda.empty_cache(). Call 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 against Linux wheels; a native Windows install hits this instead, and the mechanism is worth understanding because the code looks like it should have caught it. The graph runner's 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 what a community ComfyUI wrapper documents for official Windows wheels: "The op exists and then crashes" (ComfyUI-YuE2 README). The escape hatch is the vendor's own, not a patch: 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 above never runs. Do not pip install flash-attn in response; the package declares no such dependency and its "FlashAttention" is PyTorch's own kernel. Two caveats keep this page on 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 has been measured by nobody, and neither has its speed. Requesting flash explicitly at least 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. It is also why this recipe's floor is 16 GB rather than 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 on its 24 GB reference card, with context not reducible and the two candidate levers (--budget 12 --offload-ar) measured by nobody. A card with no documented headroom is not the same statement as a proven failure, and that distinction is exactly why 12 GB stays out until someone runs it.

Raising cfg_scale costs a second KV branch

The model card's tuning table suggests cfg_scale=1.2 for stronger 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, per the derivation above. On 16 GB that is affordable, and the two-branch case is not a guess: 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 the 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

There is 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). At 12.0 an RTX 5070 Ti clears the test with room to spare. Leave it off anyway: the module's own status output reports its quality as 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 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 that port landed on the 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 sub-9 GB peak and a real-time factor between 0.23 and 0.28, driven through that engine's interface on a 32 GB card, noting himself that "The numbers are a bit noisy because of background workloads" (vendor issue #163). Before that sub-9 GB number changes your plans, two things: it is a different engine on 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 a more aggressive cache 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 — while 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 and not 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 16 GB card of any model is named — every consumer card that appears belongs to the 24 GB tier or above. Every hardware-shaped issue you will find in the vendor's 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 carry 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.