Add cuda graphs sizes and make it default. (#1703)

# What does this PR do?

```
text-generation-launcher --model-id XXX # Uses cuda graphs by default
text-generation-launcher --model-id XXX --cuda-graphs "1,2"  #Restrict the number of cuda graphs which saves VRAM
text-generation-launcher --model-id XXX --cuda-graphs "0"  # Disabling it entirely
```
<!--
Congratulations! You've made it this far! You're not quite done yet
though.

Once merged, your PR is going to appear in the release notes with the
title you set, so make sure it's a great title that fully reflects the
extent of your awesome contribution.

Then, please replace this with a description of the change and which
issue is fixed (if applicable). Please also include relevant motivation
and context. List any dependencies (if any) that are required for this
change.

Once you're done, someone will review your PR shortly (see the section
"Who can review?" below to tag some potential reviewers). They may
suggest changes to make the code even better. If no one reviewed your PR
after a week has passed, don't hesitate to post a new comment
@-mentioning the same persons---sometimes notifications get lost.
-->

<!-- Remove if not applicable -->

Fixes # (issue)


## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
      Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
      to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
- [ ] Did you write any new necessary tests?


## Who can review?

Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.

<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @


@OlivierDehaene OR @Narsil

 -->

docs/source/basic_tutorials/launcher.md

@@ -206,12 +206,13 @@ Options:
           [env: MAX_BATCH_SIZE=]
 
 ```
-## ENABLE_CUDA_GRAPHS
+## CUDA_GRAPHS
 ```shell
-      --enable-cuda-graphs
+      --cuda-graphs <CUDA_GRAPHS>
-          Enable experimental support for cuda graphs
+          Specify the batch sizes to compute cuda graphs for. Use "0" to disable
           
-          [env: ENABLE_CUDA_GRAPHS=]
+          [env: CUDA_GRAPHS=]
+          [default: 1,2,4,8,16,32,64,96,128]
 
 ```
 ## HOSTNAME

integration-tests/conftest.py

@@ -383,7 +383,6 @@ def launcher(event_loop):
 
         env = {
             "LOG_LEVEL": "info,text_generation_router=debug",
-            "ENABLE_CUDA_GRAPHS": "true",
         }
         if not use_flash_attention:
             env["USE_FLASH_ATTENTION"] = "false"

launcher/src/main.rs

@@ -284,9 +284,15 @@ struct Args {
     #[clap(long, env)]
     max_batch_size: Option<usize>,
 
-    /// Enable experimental support for cuda graphs
+    /// Specify the batch sizes to compute cuda graphs for.
-    #[clap(long, env)]
+    /// Use "0" to disable.
-    enable_cuda_graphs: bool,
+    #[clap(
+        long,
+        env,
+        value_delimiter = ',',
+        default_value = "1,2,4,8,16,32,64,96,128"
+    )]
+    cuda_graphs: Vec<usize>,
 
     /// The IP address to listen on
     #[clap(default_value = "0.0.0.0", long, env)]
@@ -416,7 +422,7 @@ fn shard_manager(
     disable_custom_kernels: bool,
     watermark_gamma: Option<f32>,
     watermark_delta: Option<f32>,
-    enable_cuda_graphs: bool,
+    cuda_graphs: Vec<usize>,
     cuda_memory_fraction: f32,
     rope_scaling: Option<RopeScaling>,
     rope_factor: Option<f32>,
@@ -549,8 +555,16 @@ fn shard_manager(
     };
 
     // Enable experimental support for cuda graphs
-    if enable_cuda_graphs {
+    if !cuda_graphs.is_empty() {
-        envs.push(("ENABLE_CUDA_GRAPHS".into(), "True".into()))
+        envs.push((
+            "CUDA_GRAPHS".into(),
+            cuda_graphs
+                .into_iter()
+                .map(|c| c.to_string())
+                .collect::<Vec<_>>()
+                .join(",")
+                .into(),
+        ));
     }
 
     // If disable_custom_kernels is true, pass it to the shard as an env var
@@ -941,7 +955,11 @@ fn spawn_shards(
         let disable_custom_kernels = args.disable_custom_kernels;
         let watermark_gamma = args.watermark_gamma;
         let watermark_delta = args.watermark_delta;
-        let enable_cuda_graphs = args.enable_cuda_graphs;
+        let cuda_graphs: Vec<usize> = args
+            .cuda_graphs
+            .iter()
+            .filter_map(|&c| if c > 0 { Some(c) } else { None })
+            .collect();
         let cuda_memory_fraction = args.cuda_memory_fraction;
         let rope_scaling = args.rope_scaling;
         let rope_factor = args.rope_factor;
@@ -963,7 +981,7 @@ fn spawn_shards(
                 disable_custom_kernels,
                 watermark_gamma,
                 watermark_delta,
-                enable_cuda_graphs,
+                cuda_graphs,
                 cuda_memory_fraction,
                 rope_scaling,
                 rope_factor,

server/text_generation_server/models/flash_causal_lm.py

@@ -28,7 +28,7 @@ from text_generation_server.models.cache_manager import (
     BLOCK_SIZE,
 )
 from text_generation_server.pb import generate_pb2
-from text_generation_server.models.globals import MEM_POOL, ENABLE_CUDA_GRAPHS
+from text_generation_server.models.globals import MEM_POOL, CUDA_GRAPHS
 from text_generation_server.utils import StoppingCriteria, HeterogeneousNextTokenChooser
 from text_generation_server.utils.dist import MEMORY_FRACTION
 
@@ -798,11 +798,11 @@ class FlashCausalLM(Model):
             self.device,
         )
 
-        if ENABLE_CUDA_GRAPHS:
+        if CUDA_GRAPHS:
             try:
-                logger.info("Experimental support for Cuda Graphs is enabled")
+                logger.info(f"Cuda Graphs are enabled for sizes {CUDA_GRAPHS}")
                 # Warmup cuda graphs
-                for bs in [1, 2, 4] + [8 * i for i in range(1, 9)]:
+                for bs in CUDA_GRAPHS:
                     if self.speculate is None or self.speculate + 1 <= bs:
                         self.cuda_graph_warmup(bs, max_s, max_bt)
             except Exception:

server/text_generation_server/models/globals.py

@@ -3,4 +3,12 @@ import os
 
 MEM_POOL = torch.cuda.graph_pool_handle()
 # This is overridden by the cli
-ENABLE_CUDA_GRAPHS = os.getenv("ENABLE_CUDA_GRAPHS", "false").lower() in {"1", "true"}
+cuda_graphs = os.getenv("CUDA_GRAPHS")
+if cuda_graphs is not None:
+    try:
+        cuda_graphs = [int(item) for item in cuda_graphs.split(",")]
+    except Exception as e:
+        raise RuntimeError(
+            f"Could not parse cuda graphs {cuda_graphs}, expected comma separated list for batch sizes to run on: {e}"
+        )
+CUDA_GRAPHS = cuda_graphs

server/text_generation_server/models/mamba.py

@@ -13,7 +13,7 @@ from text_generation_server.utils import (
     weight_files,
     Weights,
 )
-from text_generation_server.models.globals import ENABLE_CUDA_GRAPHS, MEM_POOL
+from text_generation_server.models.globals import CUDA_GRAPHS, MEM_POOL
 import time
 from text_generation_server.models.custom_modeling.mamba_modeling import (
     MambaModel,
@@ -465,12 +465,12 @@ class Mamba(Model):
 
     def warmup(self, batch) -> Optional[int]:
         # TODO: implement warmup for Mamba if needed
-        if ENABLE_CUDA_GRAPHS:
+        if CUDA_GRAPHS:
             if self.speculate is None or self.speculate == 0:
                 try:
-                    logger.info("Experimental support for Cuda Graphs is enabled")
+                    logger.info(f"Cuda Graphs are enabled for sizes {CUDA_GRAPHS}")
                     # Warmup cuda graphs
-                    for bs in [1, 2, 4] + [8 * i for i in range(1, 9)]:
+                    for bs in CUDA_GRAPHS:
                         self.cuda_graph_warmup(bs)
                 except Exception:
                     logger.exception(f"Decode cuda graph warmup failed")