From ee92958ec0c95a3687fcc0196e9e631dd1b57b2c Mon Sep 17 00:00:00 2001 From: zanzhang Date: Tue, 20 May 2025 11:44:34 +0800 Subject: [PATCH] bugfix for varlen fmha --- .../01_fmha/codegen/ops/fmha_batch_prefill.py | 43 ++++++++++--------- ..._batch_prefill_pipeline_qr_ks_vs_async.hpp | 1 - .../ops/fmha/pipeline/tile_fmha_traits.hpp | 4 +- 3 files changed, 24 insertions(+), 24 deletions(-) diff --git a/example/ck_tile/01_fmha/codegen/ops/fmha_batch_prefill.py b/example/ck_tile/01_fmha/codegen/ops/fmha_batch_prefill.py index 57fdf098be..d20319910e 100644 --- a/example/ck_tile/01_fmha/codegen/ops/fmha_batch_prefill.py +++ b/example/ck_tile/01_fmha/codegen/ops/fmha_batch_prefill.py @@ -62,9 +62,9 @@ using fmha_trait_{F_idx} = ck_tile::TileFmhaTraits<{F_spad}, {F_lse}, {F_dropout}, {F_squant}, + {F_occupancy}, {F_sglang_layout}, - {F_chunked}, - {F_occupancy}>; + {F_chunked}>; using fmha_variant_{F_idx} = ck_tile::ComposedAttention<{F_logits} * ck_tile::LOGITS_SOFT_CAP, CK_TILE_FMHA_FWD_FAST_EXP2>; @@ -275,8 +275,8 @@ class FmhaFwdPipeline: if self.F_squant == 't' : n += '_squant' else: n += '_nsquant' - # if self.F_sglang_layout == 't' : n += '_sglang' - # else: n += '_vllm' + if self.F_sglang_layout == 't' : n += '_sglang' + else: n += '_vllm' if self.F_chunked == 't' : n += '_chunked' else: n += '_nchunked' @@ -302,6 +302,7 @@ class FmhaFwdApiPool: per_dtypes=str() for i, dtype in enumerate(self.pool.keys()): per_hdim_case=str() + for j, hdim in enumerate(self.pool[dtype].keys()): traits=self.pool[dtype][hdim] inners=str() @@ -475,31 +476,31 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, optdim_list, mask_impl squant = 't' if dtype == 'fp8' else 'f' pipelines = [] if dtype in ['fp16', 'bf16']: - for logits, mask, bias, lse, dropout, chunked in itertools.product(["t", "f"], get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["t", "f"], ["t", "f"]): + for logits, mask, bias, lse, dropout, sglang, chunked in itertools.product(["t", "f"], get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["t", "f"], ["t", "f"], ["t", "f"]): if hdim == 256: # if True: - pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, 'f', chunked)) - pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, 'f', chunked)) + pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, sglang, chunked)) + pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, sglang, chunked)) # the below two is used for hdim vectorize load - pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 'f', 'f', logits, bias, lse, dropout, squant, mask, 'f', chunked)) - pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 'f', 'f', logits, bias, lse, dropout, squant, mask, 'f', chunked)) + pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 'f', 'f', logits, bias, lse, dropout, squant, mask, sglang, chunked)) + pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 'f', 'f', logits, bias, lse, dropout, squant, mask, sglang, chunked)) - pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, 'f', chunked)) - pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, 'f', chunked)) + pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, sglang, chunked)) + pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, sglang, chunked)) else: if bias == "bias": - pipelines.append(FmhaFwdPipeline('qr_async', 'row', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, 'f', chunked)) - pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, 'f', chunked)) - pipelines.append(FmhaFwdPipeline('qr_async', 'col', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, 'f', chunked)) - pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, 'f', chunked)) + pipelines.append(FmhaFwdPipeline('qr_async', 'row', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, sglang, chunked)) + pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, sglang, chunked)) + pipelines.append(FmhaFwdPipeline('qr_async', 'col', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, sglang, chunked)) + pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, sglang, chunked)) else: - pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask, 'f', chunked)) - pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, 'f', chunked)) - pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask, 'f', chunked)) - pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, 'f', chunked)) + pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask, sglang, chunked)) + pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, sglang, chunked)) + pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask, sglang, chunked)) + pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, sglang, chunked)) if receipt == 1 and bias != "bias": - pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, 'f', chunked)) # TODO: cover arbitraty hdim - pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask, 'f', chunked)) # TODO: cover arbitraty hdim + pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, sglang, chunked)) # TODO: cover arbitraty hdim + pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask, sglang, chunked)) # TODO: cover arbitraty hdim elif dtype in ['fp8', 'bf8']: # no need lse/dropout kernels for logits, mask, bias in itertools.product(["t", "f"], get_mask_map(mask_impl).keys(), BIAS_MAP.keys()): diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_batch_prefill_pipeline_qr_ks_vs_async.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_batch_prefill_pipeline_qr_ks_vs_async.hpp index ec57af1180..0bf6c96bdf 100644 --- a/include/ck_tile/ops/fmha/pipeline/block_fmha_batch_prefill_pipeline_qr_ks_vs_async.hpp +++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_batch_prefill_pipeline_qr_ks_vs_async.hpp @@ -820,7 +820,6 @@ struct BlockFmhaBatchPrefillPipelineQRKSVSAsync } else { - statically_indexed_array k_offsets; static_for<0, NRepeat, 1>{}([&](auto n0) { int32_t seqlen_k_idx_per_repeat = k_coord[0] + kN0 / NRepeat * n0.value; int32_t i_page = seqlen_k_idx_per_repeat / page_block_size; diff --git a/include/ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp b/include/ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp index 58a8c62c65..a6e06a15b1 100644 --- a/include/ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp +++ b/include/ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp @@ -19,9 +19,9 @@ template + bool kIsChunkedPrefill_ = false> struct TileFmhaTraits { static constexpr bool kPadSeqLenQ = kPadSeqLenQ_;