[ROCm][CI] Only require q_scale==1.0 for fp8 query in RocmAttention (#46148)
Signed-off-by: stefankoncarevic <stefan.koncarevic@amd.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
co-authored by
Andreas Karatzas
parent
6ead164e52
commit
91ba720b75
@@ -421,9 +421,18 @@ class RocmAttentionImpl(AttentionImpl):
|
||||
if is_quantized_kv_cache(self.kv_cache_dtype):
|
||||
key_cache = key_cache.view(self.fp8_dtype)
|
||||
value_cache = value_cache.view(self.fp8_dtype)
|
||||
assert layer._q_scale_float == 1.0, (
|
||||
"A non 1.0 q_scale is not currently supported."
|
||||
)
|
||||
# chunked_prefill_paged_decode runs attention with a full-precision
|
||||
# query (it does not quantize Q to fp8 and does not consume
|
||||
# q_scale), so q_scale only matters when the query itself is fp8.
|
||||
# For a non-fp8 query, q_scale is not applicable and is ignored
|
||||
# (mirrors TritonAttentionImpl). This avoids spuriously failing on
|
||||
# checkpoints that carry a non-1.0 q_scale while keeping the query
|
||||
# in full precision.
|
||||
if query.dtype == self.fp8_dtype and layer._q_scale_float != 1.0:
|
||||
raise NotImplementedError(
|
||||
"A non 1.0 q_scale with an fp8 query is not currently "
|
||||
"supported by RocmAttentionImpl."
|
||||
)
|
||||
|
||||
cu_seqlens_q = attn_metadata.query_start_loc
|
||||
seqused_k = attn_metadata.seq_lens
|
||||
|
||||
Reference in New Issue
Block a user