[dtype] Map FP8 to torch.float8_e4m3fn on RDNA3

[skysnow2001]

Summary

Add gfx1100, gfx1101, gfx1102, and gfx1151 to AITER's FP8 dtype map with torch.float8_e4m3fn.

Without these entries, AITER falls back to torch.uint8. FP8 tensors produced by vLLM are then interpreted as byte tensors, which can produce invalid mixed FP8/uint8 operands in Triton kernels such as the DeepSeek-V4 paged MQA-logits path.

This follows the same dtype-map change made for gfx1200/gfx1201 in #3332.

Test

Tested on gfx1100 and gfx1151.

Test	Command	Result
Direct FP8 tgemm check	Command below	PASS — 3/3 shapes; worst max absolute difference 0.03125 on gfx1100 and 0.25 on gfx1151 with threshold 5.0
Triton contiguous FP8 MQA logits	pytest -q op_tests/triton_tests/attention/test_fp8_mqa_logits.py	PASS — 144 passed
Triton fused FP8 quant	pytest -q op_tests/triton_tests/quant/test_fused_fp8_quant.py	PASS — 393 passed
Triton fused RMS-gated FP8 group quant	pytest -q op_tests/triton_tests/quant/test_fused_rms_gated_fp8_group_quant.py	PASS — 121 passed

It passed these shapes:

4x64x128       max_abs_diff=0
16x4096x4096   max_abs_diff=0
128x1536x4096 max_abs_diff=0.03125

python3 - <<'PY'
import torch
import torch.nn.functional as F

from aiter import dtypes
from aiter.tuned_gemm import tgemm

assert dtypes.fp8 == torch.float8_e4m3fn
torch.manual_seed(0)
one = torch.ones(1, device="cuda", dtype=torch.float32)

for m, n, k in ((4, 64, 128), (16, 4096, 4096), (128, 1536, 4096)):
    x = torch.randn(m, k, device="cuda", dtype=torch.bfloat16).to(dtypes.fp8)
    w = torch.randn(n, k, device="cuda", dtype=torch.bfloat16).to(dtypes.fp8)
    ref = F.linear(x.to(torch.bfloat16), w.to(torch.bfloat16)).float()
    out = tgemm.mm(
        x, w, otype=torch.bfloat16, scale_a=one, scale_b=one
    ).float()
    max_abs_diff = (out - ref).abs().max().item()
    print(f"{m}x{n}x{k}: max_abs_diff={max_abs_diff}")
    assert max_abs_diff <= 5.0
PY

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:gfx1250-ffm-triton	Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 4764 --add-label <label>