Multi-node support for tensor parallelism (#1125)

* Multi-node support, server? * Fix check * Remove check * Remove check * Debug * Debug * counter+=1 * counter+=1 * Replicate kv heads? * Replicate kv heads? * Replicate kv heads? * Shard with specific offset * Shard with specific offset * Replicate kv heads * Fix num kv groups * Fix num kv groups * Debug * Refactor to client and server * Use multi-node synchronization * Debugging * Hierarchical synchronization design * It works! * Add some docs * Add logging * Add logging * Update docs * Try to reuse socket * Try to reuse socket * Maybe its faster * Minimize tcp traffic * Clippy * Set some timeouts
EricLBuehler · Feb 11, 2025 · 844cdc0 · 844cdc0
1 parent 3fb29cc
commit 844cdc0
Show file tree

Hide file tree

Showing 9 changed files with 622 additions and 201 deletions.
diff --git a/Cargo.lock b/Cargo.lock
diff --git a/docs/DISTRIBUTED.md b/docs/DISTRIBUTED.md
@@ -1,4 +1,4 @@
-# Distributed inference in mistral.rs
+# Distributed inference in mistral.rs: Tensor parallelism and Multi-node support
 
 Mistral.rs supports distributed inference on CUDA with Tensor Parallelism via NCCL.
 
@@ -16,3 +16,40 @@ See the following environment variables:
 |--|--|--|
 |`MISTRALRS_NO_NCCL=1`|Disable TP and NCCL|If the model does not fit on the available CUDA devices, disabling NCCL will re-enable automatic device mapping|
 |`MISTRALRS_PIPELINE_PARALLEL=<number> (default: 1 = disabled)`|Parallelize the model along the layers in addition to the GPUs|Increasing this value is useful for tuning performance on a model-specific basis. It does not change the number of GPUs required, but can help when the single-node interconnects are a bottleneck.|
+
+## Multi-node support
+
+```
+# Head node:
+MISTRALRS_MN_GLOBAL_WORLD_SIZE=32 MISTRALRS_MN_HEAD_NUM_WORKERS=1 MISTRALRS_MN_HEAD_PORT=<PORT> cargo run --release --features cuda -- -i plain -m ...
+
+# For the worker nodes:
+MISTRALRS_MN_GLOBAL_WORLD_SIZE=32 MISTRALRS_MN_WORKER_ID=0 MISTRALRS_WORKER_SERVER_ADDR=<HEAD ADDR>:<PORT> cargo run --release --features cuda -- -i plain -m ...
+MISTRALRS_MN_GLOBAL_WORLD_SIZE=32 MISTRALRS_MN_WORKER_ID=1 MISTRALRS_WORKER_SERVER_ADDR=<HEAD ADDR>:<PORT> cargo run --release --features cuda -- -i plain -m ...
+MISTRALRS_MN_GLOBAL_WORLD_SIZE=32 MISTRALRS_MN_WORKER_ID=2 MISTRALRS_WORKER_SERVER_ADDR=<HEAD ADDR>:<PORT> cargo run --release --features cuda -- -i plain -m ...
+```
+
+Multi-node support in mistral.rs divides the nodes into two groups: a "head" node, and multiple "worker" nodes. Head node choice is arbitrary.
+For example, if a system has 8 nodes, there will be 1 "head" node, and 7 "worker" nodes. 
+
+To enable multi-node, set the `MISTRALRS_MN_GLOBAL_WORLD_SIZE=<number>` environment variable to the total number of GPUs in all nodes, including "head" and "worker"s.
+
+> Note: `MISTRALRS_PIPELINE_PARALLEL` is incompatible with multi-node (setting `MISTRALRS_MN_GLOBAL_WORLD_SIZE`)
+
+It is recommended to use server mode with mistral.rs when in multi-node. **Currently, you must send requests to every node!**
+
+The following environment variables must be set for each node:
+
+**Head node:**
+
+|Name|Function|Usage|
+|--|--|--|
+|`MISTRALRS_MN_HEAD_NUM_WORKERS=<number>`|The number of worker nodes which will be connected.|This should be the number of nodes in the system, minus 1 for the head node.|
+|`MISTRALRS_MN_HEAD_PORT=<PORT>`|The port on which to communicate with the worker nodes.|Worker nodes will connect to this port via TCP sockets|
+
+**Worker node:**
+
+|Name|Function|Usage|
+|--|--|--|
+|`MISTRALRS_MN_WORKER_ID=<number>`|The 0-indexed worker ID for this worker node.|If there are 4 nodes (1 head, 3 workers), then the worker ids will be 0, 1, and 2|
+|`MISTRALRS_MN_WORKER_SERVER_ADDR=<ADDR>:<PORT>`|The IP address and port to connect to the server.|This is used to establish communication with the head node.|
diff --git a/mistralrs-core/src/models/llama.rs b/mistralrs-core/src/models/llama.rs
@@ -3,7 +3,7 @@
 use candle_core::{DType, Device, Result, Tensor};
 use candle_nn::{Embedding, Module};
 use mistralrs_quant::{
-    ColumnParallelLayer, QuantMethod, QuantizedConfig, ReplicatedLayer, RowParallelLayer,
+    ColumnParallelLayer, QuantMethod, QuantizedConfig, ReplicatedLayer, RowParallelLayer, Shard,
     ShardedVarBuilder,
 };
 use serde::{Deserialize, Serialize};
@@ -193,20 +193,39 @@ impl CausalSelfAttention {
             comm,
             vb.pp("q_proj"),
         )?;
-        let k_proj = ColumnParallelLayer::new(
+
+        // We may need to replicate the kv heads
+        let kv_replicate = if comm.world_size() > cfg.num_key_value_heads {
+            comm.world_size() / cfg.num_key_value_heads
+        } else {
+            1
+        };
+
+        let kv_shard_id = comm.rank() / kv_replicate;
+        // let kv_block_size = size_kv / comm.world_size();
+        let kv_block_size = cfg.hidden_size / cfg.num_attention_heads;
+        let shard = Shard::Offset {
+            dim: 0,
+            offset: kv_shard_id * kv_block_size,
+            len: kv_block_size,
+        };
+
+        let k_proj = ColumnParallelLayer::new_with_shard(
             size_in,
             size_kv,
             &cfg.quantization_config,
             false,
             comm,
+            shard,
             vb.pp("k_proj"),
         )?;
-        let v_proj = ColumnParallelLayer::new(
+        let v_proj = ColumnParallelLayer::new_with_shard(
             size_in,
             size_kv,
             &cfg.quantization_config,
             false,
             comm,
+            shard,
             vb.pp("v_proj"),
         )?;
         let o_proj = RowParallelLayer::new(
@@ -229,7 +248,11 @@ impl CausalSelfAttention {
             max_seq_len: cfg.max_position_embeddings,
             paged_attn,
             sdpa_params: SdpaParams {
-                n_kv_groups: cfg.num_attention_heads / cfg.num_key_value_heads,
+                n_kv_groups: if kv_replicate != 1 {
+                    (cfg.num_attention_heads / cfg.num_key_value_heads) / kv_replicate
+                } else {
+                    cfg.num_attention_heads / cfg.num_key_value_heads
+                },
                 use_flash_attn: cfg.use_flash_attn,
                 softcap: None,
                 softmax_scale: 1.0 / ((cfg.hidden_size / cfg.num_attention_heads) as f32).sqrt(),

diff --git a/mistralrs-core/src/pipeline/normal.rs b/mistralrs-core/src/pipeline/normal.rs
@@ -38,7 +38,7 @@ use crate::{
     normal_model_loader, normal_model_loader_sharded, xlora_model_loader, DeviceMapSetting,
     PagedAttentionConfig, Pipeline, Topology, TryIntoDType,
 };
-use anyhow::Result;
+use anyhow::{Context, Result};
 use candle_core::{Device, Tensor, Var};
 use hf_hub::{api::sync::ApiBuilder, Repo, RepoType};
 use indicatif::MultiProgress;
@@ -51,12 +51,12 @@ use rayon::iter::{
 use regex_automata::meta::Regex;
 use std::any::Any;
 use std::borrow::Cow;
-use std::fs;
 use std::num::{NonZero, NonZeroUsize};
 use std::path::{Path, PathBuf};
 use std::str::FromStr;
 use std::sync::{Arc, Barrier, RwLock};
 use std::time::Instant;
+use std::{env, fs};
 use tokenizers::Tokenizer;
 use tokio::sync::Mutex;
 use tracing::{info, warn};
@@ -300,7 +300,8 @@ impl Loader for NormalLoader {
 
         let use_nccl = available_devices.iter().all(|dev| dev.is_cuda())
             && available_devices.len() > 1
-            && std::env::var("MISTRALRS_NO_NCCL").is_ok_and(|x| x != "1");
+            && (std::env::var("MISTRALRS_NO_NCCL").is_err()
+                || std::env::var("MISTRALRS_NO_NCCL").is_ok_and(|x| x != "1"));
 
         // If auto, convert to Map if not using nccl
         if use_nccl {
@@ -476,15 +477,63 @@ impl Loader for NormalLoader {
                 anyhow::bail!("MISTRALRS_PIPELINE_PARALLEL must be nonzero")
             }
 
-            let world_size = available_devices.len() / pipeline_parallel_size;
+            let local_world_size = available_devices.len() / pipeline_parallel_size;
+            let global_world_size = if let Ok(x) = std::env::var("MISTRALRS_MN_GLOBAL_WORLD_SIZE") {
+                usize::from_str(&x).context("MISTRALRS_MN_GLOBAL_WORLD_SIZE")?
+            } else {
+                local_world_size
+            };
+
+            let use_multi_node = global_world_size != local_world_size;
+            if use_multi_node {
+                info!("Global world size != local world size, entering multi-node.");
+            }
 
-            info!("Tensor parallel world size is {world_size}");
+            if global_world_size < local_world_size || global_world_size % local_world_size != 0 {
+                anyhow::bail!("Global world size {global_world_size} must both be at least and divide the local world size {local_world_size}");
+            }
+
+            info!("Local tensor parallel world size is {local_world_size}");
+            info!("Global tensor parallel world size is {global_world_size}");
             info!("Pipeline parallelism size is {pipeline_parallel_size}");
 
-            let ids = (0..pipeline_parallel_size)
+            let mut ids = (0..pipeline_parallel_size)
                 .map(|_| mistralrs_quant::Id::new())
                 .collect::<Vec<_>>();
 
+            if ids.len() != 1 {
+                anyhow::bail!(
+                    "MISTRALRS_PIPELINE_PARALLEL cannot be set at the same time as MISTRALRS_MN_GLOBAL_WORLD_SIZE; multi-node is incompatible with pipeline parallel."
+                );
+            }
+
+            if use_multi_node {
+                let id = &mut ids[0];
+                if let Ok(n_nodes) = env::var("MISTRALRS_MN_HEAD_NUM_WORKERS") {
+                    let n_nodes =
+                        usize::from_str(&n_nodes).context("MISTRALRS_MN_HEAD_NUM_WORKERS")?;
+                    info!("Head node managing {n_nodes} workers.");
+                    let Ok(port) = env::var("MISTRALRS_MN_HEAD_PORT") else {
+                        anyhow::bail!(
+                            "Got MISTRALRS_MN_HEAD_NUM_WORKERS, expected MISTRALRS_MN_HEAD_PORT"
+                        );
+                    };
+                    info!("Head node initializing connection on {port}.");
+                    let server = mistralrs_quant::Server::new(
+                        &format!("0.0.0.0:{port}"),
+                        n_nodes,
+                        local_world_size,
+                    )?;
+
+                    server.broadcast_id(id)?;
+                } else if let Ok(addr) = env::var("MISTRALRS_MN_WORKER_SERVER_ADDR") {
+                    info!("Worker node connecting to {addr}.");
+                    let client = mistralrs_quant::Client::new(addr.parse()?, local_world_size)?;
+
+                    *id = client.receive_id()?;
+                }
+            }
+
             if available_devices.len() % ids.len() != 0 {
                 anyhow::bail!(
                     "Pipeline parallel size {} must divide the number of available devices {}",
@@ -497,13 +546,47 @@ impl Loader for NormalLoader {
                 .chunks(available_devices.len() / pipeline_parallel_size)
                 .collect::<Vec<_>>();
 
+            let rank_offset = if env::var("MISTRALRS_MN_WORKER_SERVER_ADDR").is_ok() {
+                let Ok(node_id) = env::var("MISTRALRS_MN_WORKER_ID") else {
+                    anyhow::bail!(
+                        "Got MISTRALRS_MN_WORKER_SERVER_ADDR, expected MISTRALRS_MN_WORKER_ID"
+                    );
+                };
+                let node_id = usize::from_str(&node_id).context("MISTRALRS_MN_WORKER_ID")?;
+                info!("Worker ID is {node_id}.");
+                (node_id + 1) * local_world_size
+            } else {
+                0
+            };
+
             // Transpose
             let mut comms_all = Vec::new();
             for (pipeline_parallel_i, devices_per_pipeline_parallel) in
                 split_available_devices.iter().enumerate()
             {
                 // Each pipeline parallel gets its own barrier
-                let barrier = Arc::new(Barrier::new(world_size));
+                let barrier = if let Ok(n_nodes) = env::var("MISTRALRS_MN_HEAD_NUM_WORKERS") {
+                    let n_nodes =
+                        usize::from_str(&n_nodes).context("MISTRALRS_MN_HEAD_NUM_WORKERS")?;
+                    let Ok(port) = env::var("MISTRALRS_MN_HEAD_PORT") else {
+                        anyhow::bail!(
+                            "Got MISTRALRS_MN_HEAD_NUM_WORKERS, expected MISTRALRS_MN_HEAD_PORT"
+                        );
+                    };
+                    let server = mistralrs_quant::Server::new(
+                        &format!("0.0.0.0:{port}"),
+                        n_nodes,
+                        local_world_size,
+                    )?;
+
+                    Arc::new(server) as Arc<dyn mistralrs_quant::BarrierLike>
+                } else if let Ok(addr) = env::var("MISTRALRS_MN_WORKER_SERVER_ADDR") {
+                    let client = mistralrs_quant::Client::new(addr.parse()?, local_world_size)?;
+                    Arc::new(client) as Arc<dyn mistralrs_quant::BarrierLike>
+                } else {
+                    Arc::new(Barrier::new(local_world_size))
+                        as Arc<dyn mistralrs_quant::BarrierLike>
+                };
 
                 // They each block on each other
                 // https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/api/comms.html?ncclcomminitrank#ncclcomminitrank
@@ -522,8 +605,8 @@ impl Loader for NormalLoader {
                         mistralrs_quant::Comm::from_device(
                             ids[pipeline_parallel_i],
                             device,
-                            rank,
-                            world_size,
+                            rank + rank_offset,
+                            global_world_size,
                             barrier.clone(),
                         )
                     })
@@ -539,7 +622,7 @@ impl Loader for NormalLoader {
 
             // row major: number of ranks x pipeline parallel
             // Also corresponds to the device for that comm for the
-            let comms = (0..world_size)
+            let comms = (0..local_world_size)
                 .map(|pipeline_parallel_i| {
                     comms_all
                         .iter()

diff --git a/mistralrs-quant/src/distributed/layers.rs b/mistralrs-quant/src/distributed/layers.rs
@@ -10,7 +10,7 @@ use crate::{
 };
 
 fn shard(dim: usize, rank: usize, world_size: usize) -> Shard {
-    Shard {
+    Shard::Simple {
         dim,
         rank,
         world_size,
@@ -195,18 +195,15 @@ pub struct ColumnParallelLayer {
 
 impl ColumnParallelLayer {
     #[allow(clippy::new_ret_no_self)]
-    pub fn new(
+    pub fn new_with_shard(
         in_dim: usize,
         out_dim: usize,
         config: &Option<QuantizedConfig>,
         bias: bool,
         comm: &Arc<crate::Comm>,
+        shard: Shard,
         vb: ShardedVarBuilder,
     ) -> Result<Arc<dyn QuantMethod>> {
-        let rank = comm.rank();
-        let world_size = comm.world_size();
-        let shard = shard(0, rank, world_size);
-
         let weight = if let Some(quant_conf) = &config {
             // GPTQ and BNB do not support tensor parallelism
             if matches!(
@@ -259,6 +256,22 @@ impl ColumnParallelLayer {
 
         Ok(Arc::new(Self { weight, bias }))
     }
+
+    #[allow(clippy::new_ret_no_self)]
+    pub fn new(
+        in_dim: usize,
+        out_dim: usize,
+        config: &Option<QuantizedConfig>,
+        bias: bool,
+        comm: &Arc<crate::Comm>,
+        vb: ShardedVarBuilder,
+    ) -> Result<Arc<dyn QuantMethod>> {
+        let rank = comm.rank();
+        let world_size = comm.world_size();
+        let shard = shard(0, rank, world_size);
+
+        Self::new_with_shard(in_dim, out_dim, config, bias, comm, shard, vb)
+    }
 }
 
 impl QuantMethod for ColumnParallelLayer {