wip fa2 triton & fix cudagraph bug

launcher/src/main.rs

@@ -1379,6 +1379,7 @@ fn main() -> Result<(), LauncherError> {
 
     let cuda_graphs = match (&args.cuda_graphs, &args.quantize) {
         (Some(cuda_graphs), Some(_q)) => cuda_graphs.clone(),
+        (Some(cuda_graphs), None) => cuda_graphs.clone(),
         #[allow(deprecated)]
         (
             None,

server/text_generation_server/models/flash_causal_lm.py

@@ -13,6 +13,7 @@ from opentelemetry import trace
 from transformers import PreTrainedTokenizerBase
 from typing import Optional, Tuple, List, Type, Dict
 
+from text_generation_server.utils.import_utils import IS_ROCM_SYSTEM
 from text_generation_server.models import Model
 from text_generation_server.utils.tokens import batch_top_tokens
 from text_generation_server.utils.speculate import get_speculate
@@ -807,7 +808,6 @@ class FlashCausalLM(Model):
             self.device,
         )
 
-        logger.info("CUDA_GRAPHS", CUDA_GRAPHS)
         if CUDA_GRAPHS:
             try:
                 logger.info(f"Cuda Graphs are enabled for sizes {CUDA_GRAPHS}")
@@ -817,8 +817,11 @@ class FlashCausalLM(Model):
                         self.cuda_graph_warmup(bs, max_s, max_bt)
             except torch.cuda.OutOfMemoryError:
                 logger.exception(f"Decode cuda graph warmup failed")
+        else:
+            logger.info(f"Cuda Graphs are disabled (CUDA_GRAPHS={CUDA_GRAPHS}).")
 
-        if IS_ROCM_SYSTEM and TUNABLEOP:
+        # TODO: fix
+        if IS_ROCM_SYSTEM and False:
             total_seqlens = list(range(16))
             for seqlen in total_seqlens:
                 self.tunableop_warmup(seqlen, max_s, max_bt)

server/text_generation_server/models/globals.py

@@ -4,11 +4,14 @@ import os
 MEM_POOL = torch.cuda.graph_pool_handle()
 # This is overridden by the cli
 cuda_graphs = os.getenv("CUDA_GRAPHS")
-if cuda_graphs is not None:
+if cuda_graphs is not None and cuda_graphs != "0":
     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}"
         )
+else:
+    cuda_graphs = None
+
 CUDA_GRAPHS = cuda_graphs

server/text_generation_server/models/mamba.py

@@ -474,6 +474,8 @@ class Mamba(Model):
                         self.cuda_graph_warmup(bs)
                 except Exception:
                     logger.exception(f"Decode cuda graph warmup failed")
+        else:
+            logger.info(f"Cuda Graphs are disabled (CUDA_GRAPHS='{CUDA_GRAPHS}').")
 
         return None
 

server/text_generation_server/utils/flash_attn.py

@@ -20,6 +20,18 @@ is_sm94 = major == 9 and minor == 4
 HAS_FLASH_ATTN = False
 HAS_FLASH_ATTN_V2_CUDA = False
 HAS_FLASH_ATTN_V2_ROCM = False
+
+ROCM_USE_FLASH_ATTN_V2_CK = False
+ROCM_USE_FLASH_ATTN_V2_TRITON = False
+
+if IS_ROCM_SYSTEM:
+    if os.getenv("ROCM_USE_FLASH_ATTN_V2_TRITON", "").lower() == "true":
+        ROCM_USE_FLASH_ATTN_V2_TRITON = True
+        logger.info("ROCm: using Flash Attention 2 Triton implementaion.")
+    else:
+        ROCM_USE_FLASH_ATTN_V2_CK = True
+        logger.info("ROCm: using Flash Attention 2 Composable Kernel implementaion.")
+
 try:
     try:
         import flash_attn_2_cuda
@@ -86,7 +98,7 @@ def attention(
     if window_size_left <= 0 and window_size_left != -1:
         raise ValueError("`window_size_left` must be > 0 or -1")
 
-    if HAS_FLASH_ATTN_V2_CUDA:
+    if IS_CUDA_SYSTEM and HAS_FLASH_ATTN_V2_CUDA:
         return flash_attn_2_cuda.varlen_fwd(
             q,
             k,
@@ -108,30 +120,7 @@ def attention(
             False,
             None,
         )
-    elif HAS_FLASH_ATTN_V2_ROCM:
-        if window_size_left != -1:
-            raise ValueError(
-                f"RoCm version of Flash Attention v2 does not support window attention (window_size_left != -1, got window_size_left={window_size_left})."
-            )
-
-        # RoCm flash API does not take the window_size_left and window_size_right arguments.
-        return flash_attn_2_cuda.varlen_fwd(
-            q,
-            k,
-            v,
-            out,
-            cu_seqlens,
-            cu_seqlens,
-            max_s,
-            max_s,
-            0.0,
-            softmax_scale,
-            False,
-            True,
-            False,
-            None,
-        )
-    elif HAS_FLASH_ATTN:
+    elif IS_CUDA_SYSTEM and HAS_FLASH_ATTN:
         if window_size_left != -1:
             raise NotImplementedError(
                 "window_size_left is only available with flash attn v2"
@@ -180,5 +169,30 @@ def attention(
             0,
             None,
         )
+    elif IS_ROCM_SYSTEM and HAS_FLASH_ATTN_V2_ROCM and ROCM_USE_FLASH_ATTN_V2_CK:
+        if window_size_left != -1:
+            raise ValueError(
+                f"RoCm version of Flash Attention v2 does not support window attention (window_size_left != -1, got window_size_left={window_size_left})."
+            )
 
-    raise NotImplementedError("flash attention is not installed")
+        # RoCm flash API does not take the window_size_left and window_size_right arguments.
+        return flash_attn_2_cuda.varlen_fwd(
+            q,
+            k,
+            v,
+            out,
+            cu_seqlens,
+            cu_seqlens,
+            max_s,
+            max_s,
+            0.0,
+            softmax_scale,
+            False,
+            True,
+            False,
+            None,
+        )
+    elif IS_ROCM_SYSTEM and ROCM_USE_FLASH_ATTN_V2_TRITON:
+        raise NotImplementedError("TODO")
+    else:
+        raise NotImplementedError(f"Flash attention is not installed (IS_CUDA_SYSTEM={IS_CUDA_SYSTEM}, IS_ROCM_SYSTEM={IS_ROCM_SYSTEM}, HAS_FLASH_ATTN_V2_CUDA={HAS_FLASH_ATTN_V2_CUDA}, HAS_FLASH_ATTN_V2_ROCM={HAS_FLASH_ATTN_V2_ROCM})")