A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 16:00 UTC
[Bugfix] Fix function names in test_block_fp8.py (#16033)
Signed-off-by: Bill Nell <bnell@redhat.com>
bnellnm committed
Apr 3, 2025 at 19:01 UTC
dcc56d62da910407fd9754843180d664280002ae
1 file changed
+5
-5
tests/kernels/test_block_fp8.py
+5
-5
@@ -360,7 +360,7 @@ def fp8_perm(m, idx):
360
return m[idx, ...]
361
362
363
-def test_moe_permute(a, a_s, topk_ids, num_groups, topk, block_m):
363
+def _moe_permute(a, a_s, topk_ids, num_groups, topk, block_m):
364
M, K = a.shape
365
366
sorted_token_ids, m_indices, num_pad = moe_align_block_size(
@@ -379,7 +379,7 @@ def test_moe_permute(a, a_s, topk_ids, num_groups, topk, block_m):
379
return a, a_s, m_indices, inv_perm
380
381
382
-def test_moe_unpermute(out, inv_perm, topk, K, topk_weight):
382
+def _moe_unpermute(out, inv_perm, topk, K, topk_weight):
383
M = topk_weight.shape[0]
384
out = out[inv_perm, ...]
385
tmp_out = out.view(-1, topk, K)
@@ -401,8 +401,8 @@ def deep_gemm_w8a8_block_fp8_moe(M, K, a, w1, w2, w1_s, w2_s, score, topk,
401
402
a_q, a_s = per_token_group_quant_fp8(a, block_m)
403
404
- a_q, a_s, m_indices, inv_perm = test_moe_permute(a_q, a_s, topk_ids,
405
- num_groups, topk, block_m)
404
+ a_q, a_s, m_indices, inv_perm = _moe_permute(a_q, a_s, topk_ids,
405
+ num_groups, topk, block_m)
406
407
inter_out = torch.zeros((a_q.shape[0], N * 2),
408
dtype=torch.bfloat16,
@@ -419,7 +419,7 @@ def deep_gemm_w8a8_block_fp8_moe(M, K, a, w1, w2, w1_s, w2_s, score, topk,
419
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(
420
(act_out_q, act_out_s), (w2, w2_s), out, m_indices)
421
422
- final_out = test_moe_unpermute(out, inv_perm, topk, K, topk_weight)
422
+ final_out = _moe_unpermute(out, inv_perm, topk, K, topk_weight)
423
424
return final_out
425