| --- |
| license: mit |
| pretty_name: BoLT Parallelism Configuration (64 GPUs) |
| task_categories: |
| - other |
| tags: |
| - bayesian-optimization |
| - benchmark |
| - parallelism-configuration |
| - tabular |
| size_categories: |
| - n<1K |
| --- |
| |
| # pco64 |
|
|
| Throughput of distributed LLM training as a function of the parallelism |
| configuration, on 64 GPUs across 16 hosts. |
| Tabular benchmark for the `bolt` problem `pco64`: the 780 |
| rows are the full candidate set. |
|
|
| - **Objective** `throughput_mean`, to maximise. `NaN` where the run OOMed, since |
| no throughput is observed at all -- a hidden (crash) constraint, not a bad value. |
| - **Constraint** `ran_successfully`. 560 of 780 |
| configurations run; the rest exhaust GPU memory. Feasibility is only learnable |
| by attempting the run, so it cannot be decoupled from the objective. |
| - **Auxiliary** `peak_mem` (GB, `inf` when infeasible), `peak_mem_normalised` |
| (the run's own feasibility margin, `< 0` iff feasible, a fixed pad value when |
| infeasible) and `overall_time` (wall clock in seconds, `NaN` when infeasible). |
|
|
| `peak_mem` does **not** define the constraint: feasible runs here reach |
| 79.3 GB, so no memory threshold separates the |
| classes. `peak_mem_normalised` does, by construction. |
|
|
| ## Instance context |
|
|
| Constant for every row; these describe the instance, not the search space. |
|
|
| | field | value | |
| |---|---| |
| | `num_layers` | 64 | |
| | `num_attention_heads` | 64 | |
| | `hidden_size` | 5120 | |
| | `ffn_hidden_size` | 25600 | |
| | `batch_size` | 512 | |
| | `seq_length` | 8192 | |
| | `log2_num_gpus` | 6 | |
| | `log2_num_hosts` | 4 | |
| | `log2_grad_accum_steps` | 0 | |
|
|
| ## Columns |
|
|
| The table is the raw sweep export, unchanged. Every knob halves or doubles, so |
| the sizes arrive as `log2_*` and their levels are evenly spaced. `zero_stage` is |
| the raw stage (0/2/3); `bolt` searches it as an ordinal 0/1/2. |
|
|
| | column | role | meaning | |
| |---|---|---| |
| | `log2_dp_size` | input | log2 of data-parallel size | |
| | `log2_tp_size` | input | log2 of tensor-parallel size | |
| | `log2_pp_size` | input | log2 of pipeline-parallel size | |
| | `log2_cp_size` | input | log2 of context-parallel size | |
| | `log2_sp_size` | input | log2 of sequence-parallel size | |
| | `log2_dp_bucket_size_mb` | input | log2 of DDP gradient bucket size in MB | |
| | `zero_stage` | input | ZeRO stage, one of [0, 2, 3] | |
| | `log2_num_model_chunks` | input | log2 of number of model chunks | |
| | `log2_grad_accum_steps` | context | log2 of gradient accumulation steps | |
| | `ran_successfully` | output | ground-truth constraint label | |
| | `throughput_mean` | output | objective, `NaN` when infeasible | |
| | `throughput_std` | output | spread over the timed steps, `NaN` when infeasible | |
| | `peak_mem_normalised` | output | feasibility margin, `< 0` iff feasible | |
| | `overall_time` | output | wall clock in seconds, `NaN` when infeasible | |
| | `all_time_steps` | output | per-step times, as a string list | |
| | `peak_mem` | output | peak GPU memory in GB, `inf` when infeasible | |
| | `peak_mem_per_gpu` | output | per-GPU peak memory, as a string list | |
|
|
| ## Best observed |
|
|
| `throughput_mean` = 0.27068 at `log2_dp_size`=4, `log2_tp_size`=2, `log2_pp_size`=0, `log2_cp_size`=0, `log2_sp_size`=2, `log2_dp_bucket_size_mb`=6, `zero_stage`=2, `log2_num_model_chunks`=0 |
| (peak memory 74.2 GB). |
|
|
| ## Provenance |
|
|
| `data_64depth-64gpus.csv`, converted to parquet unchanged. |
|
|
