feat: return reasonable generation and add integration test

integration-tests/models/__snapshots__/test_flash_phi35_moe/test_flash_llama_simple.json

@@ -0,0 +1,26 @@
+{
+  "choices": [
+    {
+      "finish_reason": "length",
+      "index": 0,
+      "logprobs": null,
+      "message": {
+        "content": "I'keeper services don't have real-time capabilities, however, I can guide you on how to find current weather conditions in Brooklyn, New York.\n\nTo get the most accurate",
+        "name": null,
+        "role": "assistant",
+        "tool_calls": null
+      },
+      "usage": null
+    }
+  ],
+  "created": 1725304474,
+  "id": "",
+  "model": "microsoft/Phi-3.5-MoE-instruct",
+  "object": "chat.completion",
+  "system_fingerprint": "2.2.1-dev0-native",
+  "usage": {
+    "completion_tokens": 40,
+    "prompt_tokens": 31,
+    "total_tokens": 71
+  }
+}

integration-tests/models/test_flash_phi35_moe.py

@@ -0,0 +1,42 @@
+import pytest
+
+
+@pytest.fixture(scope="module")
+def flash_llama_chat_handle(launcher):
+    with launcher(
+        "microsoft/Phi-3.5-MoE-instruct", 
+        num_shard=4,
+        cuda_graphs=[1, 2]
+    ) as handle:
+        yield handle
+
+
+@pytest.fixture(scope="module")
+async def flash_llama_chat(flash_llama_chat_handle):
+    await flash_llama_chat_handle.health(300)
+    return flash_llama_chat_handle.client
+
+
+@pytest.mark.private
+async def test_flash_llama_simple(flash_llama_chat, response_snapshot):
+    response = await flash_llama_chat.chat(
+        max_tokens=40,
+        seed=1337,
+        messages=[
+            {
+                "role": "system",
+                "content": "Youre a helpful assistant! Answer the users question best you can.",
+            },
+            {
+                "role": "user",
+                "content": "What is the weather like in Brooklyn, New York?",
+            },
+        ],
+    )
+
+    print(repr(response.choices[0].message.content))
+    assert (
+        response.choices[0].message.content
+        == "I'keeper services don't have real-time capabilities, however, I can guide you on how to find current weather conditions in Brooklyn, New York.\n\nTo get the most accurate"
+    )
+    assert response == response_snapshot

server/text_generation_server/models/custom_modeling/flash_llama_modeling.py

@@ -46,6 +46,7 @@ from text_generation_server.layers import (
 from text_generation_server.layers.rotary import PositionRotaryEmbedding
 from text_generation_server.layers.layernorm import (
     FastRMSNorm,
+    FastLayerNorm,
 )
 from text_generation_server.layers import (
     FastLinear,
@@ -456,17 +457,27 @@ class FlashLlamaLayer(nn.Module):
                 weights=weights,
             )
 
-        self.use_moe = config._name_or_path == "microsoft/Phi-3.5-MoE-instruct"
-
-        if self.use_moe:
+        if config._name_or_path == "microsoft/Phi-3.5-MoE-instruct":
             self.dense = BlockSparseMoE(f"{prefix}.block_sparse_moe", config, weights)
+            # with moe the layernorms are are not rmsnorms and they have bias
+            self.input_layernorm = FastLayerNorm.load(
+                prefix=f"{prefix}.input_layernorm",
+                weights=weights,
+                eps=config.rms_norm_eps,
+            )
+            self.post_attention_layernorm = FastLayerNorm.load(
+                prefix=f"{prefix}.post_attention_layernorm",
+                weights=weights,
+                eps=config.rms_norm_eps,
+            )
         else:
             self.dense = LlamaMLP(
-                prefix=f"{prefix}.mlp", config=config, weights=weights
+                prefix=f"{prefix}.mlp", config=config, weights=weights, index=index
             )
-
             self.input_layernorm = FastRMSNorm.load(
-            prefix=f"{prefix}.input_layernorm", weights=weights, eps=config.rms_norm_eps
+                prefix=f"{prefix}.input_layernorm",
+                weights=weights,
+                eps=config.rms_norm_eps,
             )
             self.post_attention_layernorm = FastRMSNorm.load(
                 prefix=f"{prefix}.post_attention_layernorm",