wip: debug gemma and flash

server/text_generation_server/layers/attention/cuda.py

@@ -293,6 +293,7 @@ else:
         max_s,
         softmax_scale,
         window_size_left=-1,
+        causal=None,
         softcap=None,
     ):
         if window_size_left != -1:
@@ -328,12 +329,33 @@ else:
                     .reshape(original_shape[0], -1, original_shape[2])
                 )
 
+        original_shape = q.shape
+        original_type = q.dtype
+
+        # ensure type is a float16 and that the last dimension is 128 or less
+        # as required by the flash attention kernel [flash-attention/csrc/flash_attn/fmha_api.cpp:246]
+        chunk_size = 128
+        if q.dtype == torch.bfloat16:
+            q = q.to(torch.float16)
+            k = k.to(torch.float16)
+            v = v.to(torch.float16)
+
+        # calculate the number of full chunks and the size of the last chunk
+        num_full_chunks = original_shape[-1] // chunk_size
+        last_chunk_size = original_shape[-1] % chunk_size
+
+        # preallocate the output tensor
         out = torch.empty_like(q)
+
+        # process full chunks
+        for i in range(num_full_chunks):
+            start = i * chunk_size
+            end = start + chunk_size
             flash_attn_cuda.fwd(
-            q,
+                q[..., start:end],
-            k,
+                k[..., start:end],
-            v,
+                v[..., start:end],
-            out,
+                out[..., start:end],
                 cu_seqlens,
                 cu_seqlens,
                 max_s,
@@ -346,4 +368,30 @@ else:
                 0,
                 None,
             )
+
+        # process the last chunk if it exists
+        if last_chunk_size > 0:
+            start = num_full_chunks * chunk_size
+            flash_attn_cuda.fwd(
+                q[..., start:],
+                k[..., start:],
+                v[..., start:],
+                out[..., start:],
+                cu_seqlens,
+                cu_seqlens,
+                max_s,
+                max_s,
+                0.0,
+                softmax_scale,
+                False,
+                True,
+                False,
+                0,
+                None,
+            )
+
+        # make sure it's the original type
+        if out.dtype != original_type:
+            out = out.to(original_type)
+
         return out

server/text_generation_server/models/__init__.py

@@ -500,11 +500,12 @@ def get_model(
     if max_input_tokens is not None and max_input_tokens <= sliding_window:
         sliding_window = -1
 
-    if (
+    should_use_sliding_window = (
-        (sliding_window is not None and sliding_window != -1)
+        sliding_window is not None and sliding_window != -1 and SUPPORTS_WINDOWING
-        and not SUPPORTS_WINDOWING
+    )
-        and max_input_tokens > sliding_window
+
-    ):
+    if should_use_sliding_window:
+        if max_input_tokens is not None and max_input_tokens > sliding_window:
             raise ValueError(
                 f"The backend {SYSTEM} does not support sliding window attention that is used by the model type {model_type}. To use this model nonetheless with the {SYSTEM} backend, please launch TGI with the argument `--max-input-tokens` smaller than sliding_window={sliding_window} (got here max_input_tokens={max_input_tokens})."
             )