Add parallelism_auto flag to automatically set dp, tp and micro batch size

architectures/centralized/client/src/app.rs

@@ -136,6 +136,7 @@ pub async fn build_app(
     .await?;
 
     let state_options: RunInitConfig<ClientId, ClientId> = RunInitConfig {
+        parallelism_auto: p.parallelism_auto,
         data_parallelism: p.data_parallelism,
         tensor_parallelism: p.tensor_parallelism,
         micro_batch_size: p.micro_batch_size,

architectures/decentralized/solana-client/src/app.rs

@@ -119,6 +119,7 @@ pub async fn build_app(
 
     let state_options: RunInitConfig<psyche_solana_coordinator::ClientId, NetworkIdentity> =
         RunInitConfig {
+            parallelism_auto: p.parallelism_auto,
             data_parallelism: p.data_parallelism,
             tensor_parallelism: p.tensor_parallelism,
             micro_batch_size: p.micro_batch_size,

nix/lib.nix

@@ -22,7 +22,8 @@ let
     || (builtins.match ".*tests/fixtures/.*$" path != null)
     || (builtins.match ".*.config/.*$" path != null)
     || (builtins.match ".*local-dev-keypair.json$" path != null)
-    || (builtins.match ".*shared/client/src/state/prompt_texts/index\\.json$" path != null);
+    || (builtins.match ".*shared/client/src/state/prompt_texts/index\\.json$" path != null)
+    || (builtins.match ".*shared/client/src/parallelism_data\\.json$" path != null);
 
   src = lib.cleanSourceWith {
     src = ../.;

psyche-book/src/enduser/create-run.md

@@ -86,11 +86,37 @@ run-manager create-run \
 
 At this point, your run has been successfully created.
 
+### Adding parallelism configuration (required for --parallelism-auto)
+
+If you want clients to use `PARALLELISM_AUTO=true` for automatic configuration, you must add a `parallelism_data.json` file to your model's HuggingFace repository.
+
+```json
+{
+	"H100": {
+		"1": { "dp": 1, "tp": 1, "micro_batch_size": 4 },
+		"8": { "dp": 4, "tp": 2, "micro_batch_size": 4 }
+	},
+	"H200": {
+		"8": { "dp": 8, "tp": 1, "micro_batch_size": 8 }
+	}
+}
+```
+
+Format: `gpu_type` → `num_gpus` → config
+
+- **gpu_type**: GPU model name (e.g., "H100", "H200")
+- **num_gpus**: Number of GPUs available (e.g., "1", "8")
+- **dp**: Data parallelism
+- **tp**: Tensor parallelism
+- **micro_batch_size**: Micro batch size per GPU
+
+The config is shared via P2P when clients join a run.
+
 ### Initializing configuration
 
 Initially, the run will not have any configuration defined and will remain paused, so no clients can join yet.
 
-To set the run configuration, you’ll need to provide mostly the same parameters as when creating the run, along with the path to a `config.toml` file that follows the [run config schema](./run-config.md).
+To set the run configuration, you'll need to provide mostly the same parameters as when creating the run, along with the path to a `config.toml` file that follows the [run config schema](./run-config.md).
 
 ```bash
 run-manager update-config \

psyche-book/src/enduser/join-run.md

@@ -117,19 +117,29 @@ though you might need to.
 
 **`NVIDIA_DRIVER_CAPABILITIES`** - An environment variable that the NVIDIA Container Toolkit uses to determine which compute capabilities should be provided to your container. It is recommended to set it to 'all', e.g. `NVIDIA_DRIVER_CAPABILITIES=all`.
 
+**`PARALLELISM_AUTO`** - Set to `true` to automatically detect optimal parallelism settings based on the model and your GPU hardware.
+
+- When enabled, the client will look up the best `DATA_PARALLELISM`, `TENSOR_PARALLELISM`, and `MICRO_BATCH_SIZE` values from a [built-in configuration table](https://github.com/PsycheFoundation/psyche/blob/main/shared/client/src/parallelism_data.json)
+- Your model and GPU hardware combination must be present in the table
+- This is the recommended option for most users
+- If set, manual parallelism settings below will be ignored
+
 **`DATA_PARALLELISM`** - Number of GPUs to distribute training data across.
 
 - If you have multiple GPUs, you can set this to 2, 4, etc. to speed up training
 - If you have 1 GPU, set this to `1`
+- Ignored if `PARALLELISM_AUTO=true`
 
 **`TENSOR_PARALLELISM`** - Number of GPUs to distribute the model across, this lets you train a model you can't fit on one single GPU.
 
 - If you have 1 GPU, set this to `1`
 - If your have `n` GPUs you can distribute the model across all of them by setting it to `n`.
+- Ignored if `PARALLELISM_AUTO=true`
 
 **`MICRO_BATCH_SIZE`** - Number of samples processed per GPU per training step
 
 - Set as high as your GPU memory allows
+- Ignored if `PARALLELISM_AUTO=true`
 
 **`AUTHORIZER`** - The Solana address that authorized your wallet to join this run
 

shared/client/Cargo.toml

@@ -36,6 +36,7 @@ clap.workspace = true
 sysinfo = "0.32.0"
 iroh.workspace = true
 iroh-blobs.workspace = true
+nvml-wrapper = "0.11.0"
 
 [features]
 parallelism = ["psyche-modeling/parallelism"]

shared/client/src/cli.rs

@@ -114,6 +114,10 @@ pub struct TrainArgs {
     #[clap(long, env, value_parser = parse_trim_quotes)]
     pub run_id: String,
 
+    /// Auto-detect parallelism settings from lookup table based on model and GPU count
+    #[clap(long, env)]
+    pub parallelism_auto: bool,
+
     #[clap(long, default_value_t = 1, env)]
     pub data_parallelism: usize,
 

shared/client/src/lib.rs

@@ -1,6 +1,7 @@
 mod cli;
 mod client;
 mod fetch_data;
+pub mod parallelism_lookup;
 mod protocol;
 mod state;
 mod tui;

shared/client/src/parallelism_lookup.rs

@@ -0,0 +1,89 @@
+use anyhow::Result;
+use hf_hub::{Repo, RepoType};
+use nvml_wrapper::Nvml;
+use serde::Deserialize;
+use std::collections::HashMap;
+use tracing::info;
+
+const REMOTE_CONFIG_FILENAME: &str = "parallelism_data.json";
+
+#[derive(Debug, Clone, Copy, Deserialize)]
+pub struct ParallelismConfig {
+    pub dp: usize,
+    pub tp: usize,
+    pub micro_batch_size: usize,
+}
+
+// Table format: gpu_type -> num_gpus -> config
+type Table = HashMap<String, HashMap<String, ParallelismConfig>>;
+
+/// Get GPU type from NVML (reads first visible GPU)
+fn get_gpu_type_from_nvml() -> Result<String> {
+    let nvml = Nvml::init()?;
+    let device = nvml.device_by_index(0)?;
+    Ok(device.name()?)
+}
+
+fn normalize_gpu_name(raw_name: &str) -> String {
+    let upper = raw_name.to_uppercase();
+    if upper.contains("H200") {
+        "H200".to_string()
+    } else if upper.contains("H100") {
+        "H100".to_string()
+    } else {
+        raw_name.to_string()
+    }
+}
+
+/// Try to load parallelism config JSON from the model's HuggingFace repo
+fn load_json_from_model_repo(model_repo_id: &str) -> Option<String> {
+    let token = std::env::var("HF_TOKEN").ok();
+
+    let api = hf_hub::api::sync::ApiBuilder::new()
+        .with_token(token)
+        .build()
+        .ok()?
+        .repo(Repo::new(model_repo_id.to_string(), RepoType::Model));
+
+    let path = api.get(REMOTE_CONFIG_FILENAME).ok()?;
+    std::fs::read_to_string(path).ok()
+}
+
+/// Lookup config in a table
+fn lookup_in_table(table: &Table, gpu_type: &str, num_gpus: usize) -> Option<ParallelismConfig> {
+    table
+        .get(gpu_type)
+        .and_then(|n| n.get(&num_gpus.to_string()))
+        .copied()
+}
+
+/// Lookup parallelism config from the model's HuggingFace repo
+pub fn lookup(model_repo_id: &str) -> Result<ParallelismConfig> {
+    let device_count = tch::Cuda::device_count() as usize;
+    if device_count == 0 {
+        anyhow::bail!("No GPUs found!");
+    }
+
+    // Use NVML for GPU type detection
+    let gpu_type = normalize_gpu_name(&get_gpu_type_from_nvml()?);
+    info!("Detected {} x {} GPU(s)", device_count, gpu_type);
+
+    let raw_json = load_json_from_model_repo(model_repo_id).ok_or_else(|| {
+        anyhow::anyhow!(
+            "No parallelism_data.json found in model repo '{}'. \
+             Add this file to use --parallelism-auto",
+            model_repo_id
+        )
+    })?;
+
+    let table: Table = serde_json::from_str(&raw_json)
+        .map_err(|e| anyhow::anyhow!("Failed to parse parallelism_data.json: {}", e))?;
+
+    info!(
+        "Using parallelism config from model repo '{}'",
+        model_repo_id
+    );
+
+    lookup_in_table(&table, &gpu_type, device_count)
+        .ok_or_else(|| anyhow::anyhow!("No config for {} x {}", device_count, gpu_type))
+}