vLLM-Ext: Resolved ALiBI bias regression

Changes:
- Optimized ALiBI memory usage.
  - Added environment variable "VLLM_PROMPT_ALIBI_MAX_SEQ_LEN" to allow
    large models to run with restricted prompt lengths.
- Added environment variable "VLLM_ALIBI_USE_FLOAT32_BIASES" to resolve
  accuracy issue on long sequences.
- Updated jais, mpt, falcon, baichuan, and bloom to work with ALiBI.
  - Due to bloom's 176B parameter size I was unable to test this model.
    Its changes are the simplest though.
- Works in lazy and eager mode.
- ALiBI is restricted to "VLLM_PROMPT_USE_FUSEDSDPA=false",
  "VLLM_CONTIGUOUS_PA=false", "VLLM_PA_SOFTMAX_IMPL=wsum_head_amax".

Reamining TODO:
- Resolve quality issue when running prompts of significantly different
  lengths.
- Resolve issue with contiguous PA.
- Integrate support for GQA along with MHA.

Co-authored-by: Tanner Voas <[email protected]>
Co-authored-by: Haihao Xiang <[email protected]>
Signed-off-by: Tanner Voas <[email protected]>

vllm_hpu_extension/ops.py

@@ -111,7 +111,8 @@ def scatter_reduce(attn, batch_size, block_groups, **rest):
 
 
 DEFAULT_PA_SOFTMAX_IMPL = "index_reduce" if "index_reduce" in capabilities() else "wsum_head_amax"
-normalize = SoftmaxNormalization(os.environ.get('VLLM_PA_SOFTMAX_IMPL', DEFAULT_PA_SOFTMAX_IMPL).split(','))
+ACTUAL_PA_SOFTMAX_IMPL = os.environ.get('VLLM_PA_SOFTMAX_IMPL', DEFAULT_PA_SOFTMAX_IMPL)
+normalize = SoftmaxNormalization(ACTUAL_PA_SOFTMAX_IMPL.split(','))
 
 
 def b2b_impl(tensor, block_mapping, matmul_op):
@@ -141,10 +142,24 @@ def block_softmax(batch_size, attn, block_mapping, block_scales, block_groups):
     return attn
 
 
-def flat_pa(query, key_cache, value_cache, block_list, block_mapping,
-            block_bias, block_scales, block_groups, scale, matmul_qk_op,
-            matmul_av_op, batch2block_matmul_op, block2batch_matmul_op,
-            keys_fetch_func, values_fetch_func):
+def flat_pa(
+    query,
+    key_cache,
+    value_cache,
+    block_list,
+    block_mapping,
+    block_bias,
+    block_scales,
+    block_groups,
+    scale,
+    position_bias,
+    matmul_qk_op,
+    matmul_av_op,
+    batch2block_matmul_op,
+    block2batch_matmul_op,
+    keys_fetch_func,
+    values_fetch_func,
+):
     batch_size = query.size(0)
     q_heads = query.size(1)
     kv_heads = key_cache.size(2)
@@ -162,10 +177,17 @@ def flat_pa(query, key_cache, value_cache, block_list, block_mapping,
     else:
         key = key.transpose(2, 3)
 
-    attn = matmul_qk_op(query, key) + block_bias
-    attn = block_softmax(batch_size, attn, block_mapping, block_scales, block_groups)
-    attn = matmul_av_op(attn, value)
-    attn = block2batch(attn, block_mapping, block2batch_matmul_op)
+    attn = matmul_qk_op(query, key)
+    attn_orig_dtype = attn.dtype
+    if position_bias is not None:
+        attn = attn.to(dtype=position_bias.dtype)
+        attn.add_(position_bias.unsqueeze(-2)[:attn.shape[0]])
+    if block_bias is not None:
+        attn = attn + block_bias.to(dtype=attn.dtype)
+    attn = block_softmax(batch_size, attn, block_mapping.to(dtype=attn.dtype), block_scales.to(dtype=attn.dtype), block_groups)
+    attn = matmul_av_op(attn, value.to(dtype=attn.dtype))
+    attn = block2batch(attn, block_mapping.to(dtype=attn.dtype), block2batch_matmul_op)
+    attn = attn.to(dtype=attn_orig_dtype)
     attn = attn.squeeze(-2)
     if kv_heads != q_heads:
         attn = attn.flatten(1, 2)
@@ -182,6 +204,7 @@ def prompt_attention(
     key: torch.Tensor,
     value: torch.Tensor,
     attn_bias: Optional[torch.Tensor] = None,
+    position_bias: Optional[torch.Tensor] = None,
     p: float = 0.0,
     scale: Optional[float] = None,
     matmul_qk_op=torch.matmul,
@@ -199,13 +222,20 @@ def prompt_attention(
             query = query.unflatten(1, (kv_heads, -1))
             key = key.unflatten(1, (kv_heads, 1))
             value = value.unflatten(1, (kv_heads, 1))
+            if position_bias is not None:
+                position_bias = position_bias.unsqueeze(2)
             if attn_bias is not None:
                 attn_bias = attn_bias.unsqueeze(2)
         attn_weights = matmul_qk_op(query * scale, key.transpose(-1, -2))
+        attn_orig_dtype = attn_weights.dtype
+        if position_bias is not None:
+            attn_weights = attn_weights.to(dtype=position_bias.dtype)
+            attn_weights.add_(position_bias)
         if attn_bias is not None:
-            attn_weights.add_(attn_bias)
+            attn_weights.add_(attn_bias.to(dtype=attn_weights.dtype))
         attn_weights = softmax_op(attn_weights, dim=-1)
-        attn_weights = matmul_av_op(attn_weights, value)
+        attn_weights = matmul_av_op(attn_weights, value.to(dtype=attn_weights.dtype))
+        attn_weights = attn_weights.to(dtype=attn_orig_dtype)
         if query_heads != kv_heads:
             attn_weights = attn_weights.flatten(1, 2)
     else: