Llama 8b f16 regression tests using random weights and inputs (#20487)

This PR adds regression tests for llama 8b f16 prefill and decode using
randomized llama 8b f16 weights and inputs. We use random llama 8b f16
weights in order to download a public version of the weights (the
original weights are gated and require access from the model authors
from the huggingface repo).

---------

Signed-off-by: aviator19941 <[email protected]>

Author: aviator19941

.github/workflows/pkgci_test_sharktank.yml

@@ -43,7 +43,7 @@ jobs:
           - name: amdgpu_rocm_mi250_gfx90a
             rocm-chip: gfx90a
             backend: rocm
-            iree_test_files: /home/esaimana/iree_tests_cache
+            iree_test_files: /groups/aig_sharks/iree-tests-cache
             sku: mi250
             target: target_hip
             runs-on: nodai-amdgpu-mi250-x86-64

tests/external/iree-test-suites/sharktank_models/benchmarks/llama/8b_f16_decode_rocm.json

@@ -0,0 +1,22 @@
+{
+    "inputs": [
+        "4x1xi64",
+        "4xi64",
+        "4xi64",
+        "4x5xi64",
+        "34x2097152xf16"
+    ],
+    "function_run": "decode_bs4",
+    "benchmark_flags": [
+        "--hip_use_streams=true",
+        "--benchmark_repetitions=10",
+        "--benchmark_min_warmup_time=3.0"
+    ],
+    "device": "hip",
+    "golden_time_tolerance_multiplier": {
+        "mi300": 1.1
+    },
+    "golden_time_ms": {
+        "mi300": 15.7
+    }
+}

tests/external/iree-test-suites/sharktank_models/benchmarks/llama/8b_f16_prefill_rocm.json

@@ -0,0 +1,21 @@
+{
+    "inputs": [
+        "4x128xi64",
+        "4xi64",
+        "4x4xi64",
+        "34x2097152xf16"
+    ],
+    "function_run": "prefill_bs4",
+    "benchmark_flags": [
+        "--hip_use_streams=true",
+        "--benchmark_repetitions=10",
+        "--benchmark_min_warmup_time=3.0"
+    ],
+    "device": "hip",
+    "golden_time_tolerance_multiplier": {
+        "mi300": 1.1
+    },
+    "golden_time_ms": {
+        "mi300": 43.1
+    }
+}

tests/external/iree-test-suites/sharktank_models/quality_tests/llama/8b_f16_decode_rocm.json

@@ -0,0 +1,30 @@
+{
+    "inputs": [
+        {
+            "source": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/toy_llama_inputs/decode_next_tokens.npy"
+        },
+        {
+            "source": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/toy_llama_inputs/decode_seq_lens.npy"
+        },
+        {
+            "source": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/toy_llama_inputs/decode_start_positions.npy"
+        },
+        {
+            "source": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/toy_llama_inputs/decode_seq_block_ids.npy"
+        },
+        {
+            "source": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/toy_llama_inputs/decode_cache_state.npy"
+        }
+    ],
+    "device": "hip",
+    "real_weights": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/real_weights.irpa",
+    "mlir": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/8b_f16_random.mlir",
+    "compiler_flags": [
+        "--iree-hal-target-device=hip",
+        "--iree-opt-level=O3",
+        "--iree-hal-indirect-command-buffers=true",
+        "--iree-stream-resource-memory-model=discrete",
+        "--iree-hal-memoization=true"
+    ],
+    "run_function": "decode_bs4"
+}

tests/external/iree-test-suites/sharktank_models/quality_tests/llama/8b_f16_prefill_rocm.json

@@ -0,0 +1,27 @@
+{
+    "inputs": [
+        {
+            "source": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/toy_llama_inputs/prefill_token_ids.npy"
+        },
+        {
+            "source": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/toy_llama_inputs/prefill_seq_lens.npy"
+        },
+        {
+            "source": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/toy_llama_inputs/prefill_seq_block_ids.npy"
+        },
+        {
+            "source": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/toy_llama_inputs/prefill_cache_state.npy"
+        }
+    ],
+    "device": "hip",
+    "real_weights": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/real_weights.irpa",
+    "mlir": "https://sharkpublic.blob.core.windows.net/sharkpublic/halo-models/llm-dev/llama3_8b_random/8b_f16_random.mlir",
+    "compiler_flags": [
+        "--iree-hal-target-device=hip",
+        "--iree-opt-level=O3",
+        "--iree-hal-indirect-command-buffers=true",
+        "--iree-stream-resource-memory-model=discrete",
+        "--iree-hal-memoization=true"
+    ],
+    "run_function": "prefill_bs4"
+}