[Perf] Add do_not_specialize to Mamba SSD chunk kernels (#43803)

Signed-off-by: Majid Taheri Andani <tahemaji@amazon.com>
Co-authored-by: Majid Taheri Andani <tahemaji@amazon.com>
This commit is contained in:
Majid
2026-05-28 15:40:02 +00:00
committed by GitHub
co-authored by Majid Taheri Andani
parent be4062fd6c
commit c08ebebf30
4 changed files with 0 additions and 10 deletions
@@ -33,7 +33,6 @@ def _causal_conv1d_fwd_kernel( # continuous batching
o_ptr, # (dim, seqlen) - actually pointing to x_ptr o_ptr, # (dim, seqlen) - actually pointing to x_ptr
# Matrix dimensions # Matrix dimensions
dim: tl.constexpr, dim: tl.constexpr,
seqlen: tl.int32, # cu_seqlen
num_cache_lines: tl.constexpr, # added to support vLLM larger cache lines num_cache_lines: tl.constexpr, # added to support vLLM larger cache lines
# Strides # Strides
stride_x_dim: tl.constexpr, # stride to get to next feature-value, stride_x_dim: tl.constexpr, # stride to get to next feature-value,
@@ -715,7 +714,6 @@ def causal_conv1d_fn(
out, out,
# Matrix dimensions # Matrix dimensions
dim, dim,
cu_seqlen,
num_cache_lines, num_cache_lines,
# stride # stride
stride_x_dim, stride_x_dim,
@@ -69,7 +69,6 @@ def _bmm_chunk_fwd_kernel(
out_ptr, out_ptr,
cu_chunk_seqlens_ptr, cu_chunk_seqlens_ptr,
# Matrix dimensions # Matrix dimensions
seqlen,
chunk_size: tl.constexpr, chunk_size: tl.constexpr,
K: tl.constexpr, K: tl.constexpr,
ngroups: tl.constexpr, ngroups: tl.constexpr,
@@ -191,7 +190,6 @@ def _bmm_chunk_fwd(a, b, chunk_size, cu_chunk_seqlens, causal=False, output_dtyp
b_ptr=b, b_ptr=b,
out_ptr=out, out_ptr=out,
cu_chunk_seqlens_ptr=cu_chunk_seqlens, cu_chunk_seqlens_ptr=cu_chunk_seqlens,
seqlen=seqlen,
chunk_size=chunk_size, chunk_size=chunk_size,
K=k, K=k,
ngroups=ngroups, ngroups=ngroups,
@@ -163,7 +163,6 @@ def _chunk_scan_fwd_kernel(
chunk_size: tl.constexpr, chunk_size: tl.constexpr,
hdim: tl.constexpr, hdim: tl.constexpr,
dstate: tl.constexpr, dstate: tl.constexpr,
seqlen,
nheads_ngroups_ratio: tl.constexpr, nheads_ngroups_ratio: tl.constexpr,
# Strides # Strides
stride_cb_chunk: tl.int64, stride_cb_chunk: tl.int64,
@@ -482,7 +481,6 @@ def _chunk_scan_fwd(
chunk_size=chunk_size, chunk_size=chunk_size,
hdim=headdim, hdim=headdim,
dstate=dstate, dstate=dstate,
seqlen=seqlen,
nheads_ngroups_ratio=nheads // ngroups, nheads_ngroups_ratio=nheads // ngroups,
stride_cb_chunk=cb.stride(0), stride_cb_chunk=cb.stride(0),
stride_cb_head=cb.stride(1), stride_cb_head=cb.stride(1),
@@ -35,7 +35,6 @@ def _chunk_cumsum_fwd_kernel(
dA_cumsum_ptr, dA_cumsum_ptr,
cu_chunk_seqlens_ptr, cu_chunk_seqlens_ptr,
# Matrix dimension # Matrix dimension
seqlen,
nheads: tl.constexpr, nheads: tl.constexpr,
chunk_size: tl.constexpr, chunk_size: tl.constexpr,
dt_min: tl.constexpr, dt_min: tl.constexpr,
@@ -206,7 +205,6 @@ def _chunk_state_fwd_kernel(
hdim: tl.constexpr, hdim: tl.constexpr,
dstate: tl.constexpr, dstate: tl.constexpr,
chunk_size: tl.constexpr, chunk_size: tl.constexpr,
seqlen,
nheads_ngroups_ratio: tl.constexpr, nheads_ngroups_ratio: tl.constexpr,
# Strides # Strides
stride_x_seqlen: tl.int64, stride_x_seqlen: tl.int64,
@@ -331,7 +329,6 @@ def _chunk_cumsum_fwd(
dt_out_ptr=dt_out, dt_out_ptr=dt_out,
dA_cumsum_ptr=dA_cumsum, dA_cumsum_ptr=dA_cumsum,
cu_chunk_seqlens_ptr=cu_chunk_seqlens, cu_chunk_seqlens_ptr=cu_chunk_seqlens,
seqlen=seqlen,
nheads=nheads, nheads=nheads,
chunk_size=chunk_size, chunk_size=chunk_size,
dt_min=dt_limit[0], dt_min=dt_limit[0],
@@ -389,7 +386,6 @@ def _chunk_state_fwd(
hdim=headdim, hdim=headdim,
dstate=dstate, dstate=dstate,
chunk_size=chunk_size, chunk_size=chunk_size,
seqlen=seqlen,
nheads_ngroups_ratio=nheads // ngroups, nheads_ngroups_ratio=nheads // ngroups,
stride_x_seqlen=x.stride(0), stride_x_seqlen=x.stride(0),
stride_x_head=x.stride(1), stride_x_head=x.stride(1),