[BUG] Multi-GPU tensor-split offloads all physical layers to GPU0 on looped architectures

Aperta
#206 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
3/5
Tempo stimato
Mezza giornata
Idoneità per principianti
58/100
Tipo di issue
Bug
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
c, cpp
Ambito
backend

Direzione di ricerca

Start in src/llama-model.cpp at llama_model_base::create_tensor, where tn.bid (the GGUF physical block id) indexes pimpl->dev_layer, which is sized to n_layer_all; read loop_visit and loop_layer in the hparams/model code to confirm the mapping described. Apply the proposed change so the index goes through hparams.loop_visit(tn.bid) and check that non-looped models still get identity mapping. Done means a looped model loaded with --tensor-split places postlude blocks on the second GPU and no CUDA0 OOM occurs, which needs multi-GPU hardware.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

[BUG] Multi-GPU tensor-split offloads all physical layers to GPU0 on looped architectures

Description

When loading a looped model (such as Inkling-2-Mini, where 24 physical layers expand to 48 graph execution layers) with --tensor-split across multiple GPUs, all physical layers are erroneously assigned to device 0 (CUDA0), starving secondary GPUs (CUDA1) and causing an immediate out-of-memory crash:

ggml_backend_cuda_buffer_type_alloc_buffer: allocating 11383.97 MiB on device 0: cudaMalloc failed: out of memory
ggml_backend_buft_alloc_buffer_n_default: failed to allocate CUDA0 buffer of size 11936954112

Root Cause

  1. In GGUF files for looped models, physical layer blocks are numbered blk.0 to blk.23 (n_layer_weights = 24), whereas the graph execution sequence contains n_layer_all = 48 layers:

    • 0..5: prelude (physical 0..5)
    • 6..17: loop pass 0 (physical 6..17)
    • 18..29: loop pass 1 (physical 6..17)
    • 30..41: loop pass 2 (physical 6..17)
    • 42..47: postlude (physical 18..23)
  2. In llama_model_base::create_tensor:

    if (tn.bid != -1) {
        buft_list_layer = (size_t) tn.bid < pimpl->dev_layer.size() ? pimpl->dev_layer.at(tn.bid).buft_list : pimpl->dev_output.buft_list;
    }
    

    Here tn.bid is the physical block index in the GGUF file (0..23).
    pimpl->dev_layer is sized to n_layer_all (48), where pimpl->dev_layer[il] = get_layer_buft_list(loop_layer(il)).

    When create_tensor is called for postlude physical blocks 18..23:
    It indexes dev_layer.at(tn.bid), which evaluates to dev_layer.at(18..23).
    Because graph slots 18..23 are pass 1 of the looped core, loop_layer(18..23) maps to 6..11 (GPU0).
    Consequently, physical layers 18..23 (which were intended for GPU1) receive the device assignment of physical layers 6..11 (GPU0).
    Physical layers 0..17 are also mapped to 0..17 (GPU0).
    Thus, all 24 physical layers in the file are allocated on GPU0, regardless of --tensor-split.

Proposed Fix

In src/llama-model.cpp:
Map tn.bid to its canonical graph visit slot via hparams.loop_visit(tn.bid):

ggml_tensor * llama_model_base::create_tensor(llama_model_loader & ml, const LLM_TN_IMPL & tn, const std::initializer_list<int64_t> & ne, int flags) {
    const buft_list_t * buft_list_layer = nullptr;
    if (tn.bid != -1) {
        // blocks that are not model layers (e.g. the blocks of a head) go with the output
        const size_t iv = (size_t) hparams.loop_visit(tn.bid);
        buft_list_layer = iv < pimpl->dev_layer.size() ? pimpl->dev_layer.at(iv).buft_list : pimpl->dev_output.buft_list;
    }
    return ml.create_tensor(
        hparams, &pimpl->cpu_buft_list, pimpl->dev_input.buft_list, pimpl->dev_output.buft_list, buft_list_layer,
        tn, ne, flags);
}

For non-looped models, loop_visit(tn.bid) == tn.bid, preserving existing behavior.
For looped models, loop_visit(18..23) yields 42..47, correctly retrieving the device placement assigned to the postlude layers.

Lingua principale
C++
Stelle
1.2k
Fork
80
Merge medio
4g 20h
PR unite (30g)
2

Preparare l'ambiente

Questo progetto non fornisce container di sviluppo, Dockerfile né guida per i contributori, quindi l'ambiente è a tuo carico: parti dal suo README e consulta la nostra guida al primo contributo per i passaggi generali.

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di Anbeeld/beellama.cpp

Tutte le issue di Anbeeld/beellama.cpp

Issue simili

Altre issue su C++

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.