docker: use FA3 3.0.0 for transformer_engine 2.17

[yueming-yuan]

Problem

#1781 bumped transformer_engine to 2.17.0. TE 2.17 moved use_flash_attn_3 out
of the window_size-tuple branch in cp_p2p_fwd_flash_attn /
cp_p2p_bwd_flash_attn:

# TE 2.12
if use_flash_attn_3 or (v2_3_plus and not v2_7_0_plus):
    fa_forward_kwargs["window_size"] = (-1, -1)        # FA3 took the tuple
elif v2_7_0_plus:
    fa_forward_kwargs["window_size_left"] = -1

# TE 2.17
if v2_3_plus and not v2_7_0_plus:
    fa_forward_kwargs["window_size"] = (-1, -1)
elif use_flash_attn_3 or v2_7_0_plus:                  # FA3 moved here
    fa_forward_kwargs["window_size_left"] = -1

The flash_attn_3 wheel we ship was 3.0.0b1, whose _flash_attn_forward only
accepts a single window_size tuple, so every context-parallel path raises
TypeError: got an unexpected keyword argument 'window_size_left'. There are
eight such call sites in context_parallel.py (five forward, three backward).

TE has no FA3-only kill switch — NVTE_FLASH_ATTN disables FA2 and FA3
together. FA2 in the same image (2.7.4.post1) already takes the new kwargs, so
only the FA3 path is affected.

Why this never broke before

TE's version gate is a silent capability gate that also guards the import. Under
TE 2.12 the tuple was correct, so nothing failed. The 2.17 bump is the first
time this pairing is exercised — and main's scheduled docker builds have been
failing since #1781 landed, so :dev still carries TE 2.12 and main looks green.

Change

yueming-yuan/miles-wheels#10 rebuilt the wheel at flash-attention 00756db
(version 3.0.0) and replaced the cu130-x86_64 release asset, so the existing
pip install /tmp/wheels/flash_attn_3-*.whl already resolves to it.

This PR only drops the separate flash_attn_interface.py fetch. That fetch was a
second revision pin that can drift from the one the extension module was built
at. It is now both unnecessary and harmful:

• the wheel ships flash_attn_interface top-level, which is what TE 2.17 imports;
• it also ships flash_attn_3.flash_attn_interface as a re-export shim;
• overwriting that shim with a copy of the full module re-runs
  @torch.library.custom_op("flash_attn_3::_flash_attn_forward") and
  double-registers the ops.

Nothing in miles or sglang imports flash_attn_3.flash_attn_interface.

Verification

Built and installed on an h200-sci devbox running the miles image itself
(torch 2.11.0+cu130, CUDA 13.0), so the wheel matches the toolchain it ships with:

[INFO     | DotProductAttention]: Running with FlashAttention backend (version 3.0.0)
[PASS] all 5 symbols TE imports exist
[PASS] _flash_attn_forward schema takes window_size_left/right  ['SymInt window_size_left=-1', 'SymInt window_size_right=-1']
[PASS] _flash_attn_backward schema takes window_size_left/right  ['SymInt window_size_left=-1', 'SymInt window_size_right=-1']
[PASS] import transformer_engine.pytorch
[PASS] TE sees flash-attn-3 == 3.0.0  fa3_version=3.0.0 v3_0_0_beta=False
[PASS] fwd accepts window_size_left/right kwargs  out=(2, 512, 8, 128) lse=(2, 8, 512)
[PASS] returns (out, softmax_lse, ...) in the order TE indexes [0]/[1]
[PASS] TE DotProductAttention fwd+bwd on FlashAttention backend  grad_finite=True
[PASS] output matches SDPA reference  max_abs_err=0.001953
RESULT: ALL PASS

Checked so it is not rediscovered later:

• Arch coverage is unchanged: sm_80 + sm_90a, confirmed with cuobjdump
  against both the old and new _C.abi3.so. hopper/setup.py derives arch from
  source-file suffixes and ignores TORCH_CUDA_ARCH_LIST, so FA3 has never had
  Blackwell kernels here.
• 3.0.0b1 -> 3.0.0 is benign. It flips FlashAttentionUtils.v3_0_0_beta
  to False; its only consumer appends an "update your flash-attn v3 beta" hint
  to an already-raised TypeError. No functional branch depends on it.
• Return shape is compatible. 3.0.0 returns
  (out, softmax_lse, out_accum, softmax_lse_accum); TE reads [0] and [1]
  under use_flash_attn_3.

Full CP coverage runs in #1795, which cherry-picks this change.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.