Remove FlashInfer benchmark backends

.github/workflows/collectivex-sweep.yml

@@ -10,7 +10,7 @@ on:
         description: "EP library to sweep — 'all' runs every EP backend in one matrix"
         type: choice
         default: all
-        options: [all, deepep-v2, mori, uccl-ep, nccl-ep, flashinfer-ep]
+        options: [all, deepep-v2, mori, uccl-ep, nccl-ep]
       only_sku:
         description: Restrict to one GHA runner pool; blank = all
         type: string

experimental/operatorx/CLUSTERS.md

@@ -63,7 +63,7 @@ OPERATORX_PARTITION=<partition> \
 OPERATORX_ACCOUNT=<account> \
 OPERATORX_QOS=<qos> \
 OPERATORX_SQUASH_DIR=<squash-dir> \
-OPERATORX_BACKENDS=torch,deepgemm,flashinfer,sglang \
+OPERATORX_BACKENDS=torch,deepgemm,sglang \
 OPERATORX_JOB_NAME=<job-name> \
 python3 scripts/submit_run.py nvidia
 ```

experimental/operatorx/README.md

@@ -36,7 +36,7 @@ OPERATORX_PARTITION=batch_1 \
 OPERATORX_ACCOUNT=benchmark \
 OPERATORX_QOS=batch_1_qos \
 OPERATORX_SQUASH_DIR=/data/home/sa-shared/harrison/containers \
-OPERATORX_BACKENDS=torch,deepgemm,flashinfer,sglang \
+OPERATORX_BACKENDS=torch,deepgemm,sglang \
 python3 scripts/submit_run.py nvidia
 ```
 

experimental/operatorx/containers.toml

@@ -1,9 +1,6 @@
 [nvidia.torch]
 image = "nvcr.io/nvidia/pytorch:26.03-py3"
 
-[nvidia.flashinfer]
-image = "lmsysorg/sglang:latest"
-
 [nvidia.deepgemm]
 image = "lmsysorg/sglang:latest"
 
@@ -13,9 +10,6 @@ image = "lmsysorg/sglang:latest"
 [nvidia.deepep]
 image = "lmsysorg/sglang:latest"
 
-[nvidia.flashinfer_comm]
-image = "lmsysorg/sglang:latest"
-
 [nvidia.sglang_comm]
 image = "lmsysorg/sglang:latest"
 

experimental/operatorx/runners/nvidia/backends/torch.py

@@ -80,7 +80,7 @@ def _prepare_attention_mha(op: Op) -> dict:
     for k in ("dtype_q", "dtype_k", "dtype_v"):
         if a[k] == "fp8":
             raise UnsupportedOpError(
-                f"torch SDPA doesn't support fp8 ({k}={a[k]}); use flashinfer for fp8 attention"
+                f"torch SDPA doesn't support fp8 ({k}={a[k]})"
             )
     if not (a["dtype_q"] == a["dtype_k"] == a["dtype_v"]):
         raise UnsupportedOpError(
@@ -271,7 +271,7 @@ def _kernel_combine(ctx: dict) -> None:
     _kernel_dispatch(ctx)
 
 
-# moe_forward is owned by the sglang and flashinfer backends (production MoE
+# moe_forward is owned by the sglang backend (production MoE
 # kernels). The torch backend doesn't expose its own MoE module.
 
 

experimental/operatorx/runners/nvidia/runner.py

@@ -6,7 +6,7 @@
 
 from operatorx.core import BackendImpl, Op, Result, UnsupportedOpError
 
-_BACKENDS = ["torch", "deepgemm", "flashinfer", "deepep", "sglang", "flashinfer_comm", "sglang_comm"]
+_BACKENDS = ["torch", "deepgemm", "deepep", "sglang", "sglang_comm"]
 _DISPATCH: dict[tuple[str, str], BackendImpl] = {}
 _L2_BUF: dict[int, torch.Tensor] = {}