#5337: ttnn fixes

models/demos/wormhole/mistral7b/tests/test_mistral_model_prefill.py

@@ -11,6 +11,8 @@
     os.environ["MISTRAL_CKPT_DIR"] = "/mnt/MLPerf/ttnn/models/demos/mistral7b/"
     os.environ["MISTRAL_TOKENIZER_PATH"] = "/mnt/MLPerf/ttnn/models/demos/mistral7b/"
     os.environ["MISTRAL_CACHE_PATH"] = "/mnt/MLPerf/ttnn/models/demos/mistral7b/"
+    # Prefill prompt files too large to keep in repo
+    os.environ["MIXTRAL_REF_OUTPUT_PATH"] = "/mnt/MLPerf/tt_dnn-models/Mistral/Mixtral-8x7B-v0.1/prefill/"
 
 import ttnn
 from models.demos.wormhole.mistral7b.tt.mistral_common import (
@@ -45,8 +47,10 @@ def forward(self, x):
     "seq_len",
     (
         128,
-        # 1024//2,
-        # 1024 * 2,
+        512,
+        1024,
+        2048,
+        4096,
     ),
 )
 @pytest.mark.parametrize(
@@ -68,6 +72,7 @@ def test_mistral_model_inference(device, version, seq_len, use_program_cache, re
     cache_pcc = False  # Flag to measure KV cache PCC for all layers
 
     dtype = ttnn.bfloat8_b
+    pcc = 0.94
 
     model_args = TtModelArgs(device)
     model_args.n_layers = 32  # Full model
@@ -90,7 +95,11 @@ def test_mistral_model_inference(device, version, seq_len, use_program_cache, re
         # The instruct prompts follow the format: <bos> [INST] prompt [/INST]. [INST] are strings. <bos> is the correspoding bos_id token
         prompts = ["[INST] what is the capital of Canada? [/INST]"] * 32
     else:
-        prompt_file = "models/demos/wormhole/mistral7b/tests/tale-of-two-cities.txt"
+        prompt_file = os.environ["MIXTRAL_REF_OUTPUT_PATH"] + "/tale-of-two-cities.txt"
+        assert os.path.exists(
+            prompt_file
+        ), f"Expected prompt file not found: {prompt_file}. Please set the flag 'MIXTRAL_REF_OUTPUT_PATH' correctly."
+
         with open(prompt_file, "r") as f:
             prompts = f.read()
     encoded_prompts = tokenizer.encode(prompts)[:seq_len]
@@ -156,7 +165,7 @@ def test_mistral_model_inference(device, version, seq_len, use_program_cache, re
         # TODO Measure only PCC at the end, instead of at every iteration
         # Measure PCC if also running reference model
         if run_ref_pt:
-            passing, pcc_message = comp_pcc(ref_output, tt_output_torch, 0.94)
+            passing, pcc_message = comp_pcc(ref_output, tt_output_torch, pcc)
 
             logger.info(comp_allclose(ref_output, tt_output_torch))
             logger.info(f"Model output: {pcc_message}")
@@ -206,4 +215,4 @@ def test_mistral_model_inference(device, version, seq_len, use_program_cache, re
             logger.info(f"All  Mistral decode iterations Passed!")
         else:
             logger.warning("One or more iterations of Mistral decode had bad PCC")
-            assert all_tests_pass, f"PCC value is lower than {0.99} for some of the outputs. Check Warnings!"
+            assert all_tests_pass, f"PCC value is lower than {pcc} for some of the outputs. Check Warnings!"

models/demos/wormhole/mistral7b/tt/model_config.py

@@ -141,7 +141,7 @@ def __init__(self, device, instruct=False):
                 per_core_M=4,  # 32, #16,  # M / TILE_HEIGHT / Grid_Size (dynamic based on seqlen)
                 per_core_N=56,  # N / TILE_WIDTH / Grid_Size
                 transpose_mcast=False,
-                fused_activation=ttnn.experimental.tensor.FusibleActivation.SILU,
+                fused_activation=ttnn.UnaryOpType.SILU,
                 fuse_batch=False,
             )
 
@@ -176,7 +176,7 @@ def __init__(self, device, instruct=False):
                 per_core_M=1,  # 32, #16,  # M / TILE_HEIGHT / Grid_Size (dynamic based on seqlen)
                 per_core_N=56,  # N / TILE_WIDTH / Grid_Size
                 transpose_mcast=False,
-                fused_activation=ttnn.experimental.tensor.FusibleActivation.SILU,
+                fused_activation=ttnn.UnaryOpType.SILU,
                 fuse_batch=False,
             )