[CI][ROCm] Add ten AMD parity groups (#57876)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Andreas Karatzas
2026-09-21 11:08:06 +00:00
committed by GitHub
co-authored by OpenAI Codex
parent 15859bb3a1
commit 82daf9f575
16 changed files with 525 additions and 3 deletions
@@ -0,0 +1,5 @@
model_name: "Qwen/Qwen3-30B-A3B"
accuracy_threshold: 0.88
num_questions: 1319
num_fewshot: 5
server_args: "--enforce-eager --max-model-len 8192 --tensor-parallel-size 2 --moe-backend=triton"
@@ -0,0 +1 @@
Qwen3-30B-A3B-BF16-triton-rocm.yaml
@@ -0,0 +1 @@
Qwen3-30B-A3B-Fp8-AutoFp8-triton.yaml
@@ -0,0 +1 @@
Qwen3-30B-A3B-Fp8-CT-Block-triton.yaml
@@ -0,0 +1 @@
Mixtral-8x7B-Fp8-AutoFp8-triton.yaml
@@ -0,0 +1,26 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
model_name: "Qwen/Qwen3.8-Flash-Next-FP8"
datasets:
gsm8k:
metric_threshold: 0.98
tolerance: 0.02
aime25:
metric_threshold: 0.90
tolerance: 0.05
eval_batch_size: 32
startup_max_wait_seconds: 1800
generation_config:
seed: 1236
do_sample: true
temperature: 0.7
top_p: 0.8
top_k: 20
repetition_penalty: 1.0
presence_penalty: 1.5
max_tokens: 32768
stream: true
extra_body:
chat_template_kwargs:
enable_thinking: true
server_args: "--tensor-parallel-size 4 --mamba-cache-mode align"
@@ -0,0 +1 @@
Qwen3.8-Flash-Next-FP8-rocm.yaml
@@ -96,7 +96,7 @@ def apply_rope_gptj_last_k(
# `e*c - o*s` fused form. This keeps the reference close to the kernel, but
# the fp32 reference and the fp32 GPU kernel can still round to bf16 on
# opposite sides of a round-to-nearest tie for a tiny number of elements at
# high positions, so callers compare the RoPE region within 1 bf16 ULP.
# high positions, so callers allow for bf16 rounding in the RoPE region.
new_even = torch.addcmul(-odd * sin, even, cos)
new_odd = torch.addcmul(odd * cos, even, sin)
rope_rotated = torch.stack((new_even, new_odd), dim=-1).reshape(shape)
@@ -314,7 +314,19 @@ def test_q_path_without_qnorm_matches_rope_only_reference(num_tokens: int):
apply_q_norm=False,
)
torch.testing.assert_close(q_out[:, :n_heads], q_ref, rtol=0, atol=0)
torch.testing.assert_close(
q_out[:, :n_heads, :NOPE_DIM], q_ref[..., :NOPE_DIM], rtol=0, atol=0
)
# HIP and PyTorch can contract different FP32 products into an FMA.
# Allow bf16 rounding and small cancellation errors only in the RoPE region
# on ROCm; the unrotated dimensions and padded heads must remain exact.
rope_rtol, rope_atol = (1e-2, 1e-6) if current_platform.is_rocm() else (0, 0)
torch.testing.assert_close(
q_out[:, :n_heads, NOPE_DIM:],
q_ref[..., NOPE_DIM:],
rtol=rope_rtol,
atol=rope_atol,
)
assert q_out[:, n_heads:].count_nonzero().item() == 0