Load test setup from config files, enable plugin in V1 (#25)

This PR adds to major features
1. loading micro-benchmark configurations from files: 
```
python3 /scripts/benchmark.py prefix -x /scripts/setups/prefix_correctness_rocm.conf
```
However, environment variables for `MY_IUT`, `MY_METHODS`, and `TRITON_BACKEND_DEBUG`  would overwrite the values from the config files. 

2. enable the vllm-triton-backend plugin for vLLM V1 (and stop supporting vLLM V0): So, after installation of ibm-triton-lib, the plugin can be used with just:
```
vllm serve meta-llama/Llama-3.1-8B-Instruct
```
Additionally, it fixes a correctness issue with the prefix_prefill micro-benchmark and allows a finer grained composition of batches for prefix prefill. 

---------

Signed-off-by: Burkhard Ringlein <ngl@zurich.ibm.com>

Author: bringlein

Makefile

@@ -3,7 +3,7 @@ MAX_JOBS := 64
 
 SHELL := /bin/bash
 
-.PHONY: all build clean format dev rocm rocm-upstream pyupdate nightly bm-rocm
+.PHONY: all build clean format dev rocm rocm-upstream pyupdate nightly bm-rocm spelling
 
 all: build
 
@@ -65,3 +65,7 @@ else
 format:
 	python -m black --check --verbose scripts ibm-triton-lib third_party
 endif
+
+spelling:
+	codespell ./ibm-triton-lib ./triton-dejavu ./scripts
+

ibm-triton-lib/ibm_triton_lib/backend/__init__.py

@@ -19,11 +19,4 @@
 
 def register():
     """Register the triton attention platform."""
-
-    VLLM_USE_V1 = int(os.environ.get("VLLM_USE_V1", "0"))
-
-    # backend only works with v0 currently
-    if VLLM_USE_V1:
-        return None
-    else:
-        return "ibm_triton_lib.backend.platform.TritonPlatform"
+    return "ibm_triton_lib.backend.platform.TritonPlatform"

ibm-triton-lib/ibm_triton_lib/backend/platform.py

@@ -61,4 +61,6 @@ def get_attn_backend_cls(
         use_v1,
         use_mla,
     ) -> str:
+        if not envs.VLLM_USE_V1:
+            raise RuntimeError("vllm-triton-backend plugin only supports vLLM V1")
         return "ibm_triton_lib.backend.triton_attn.TritonAttentionBackend"