Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In `@tests/utils/test_fp4_quantize.py`:
- Around line 706-747: In test_nvfp4_per_token_quantize_te_reference ensure the
FOUR_OVER_SIX mode is pinned off so the TE-reference path is deterministic: at
the start of test_nvfp4_per_token_quantize_te_reference set the environment flag
FLASHINFER_NVFP4_FOUR_OVER_SIX="0" (or call your library’s setter if available)
before creating x and running ref_fp4_quant_te/nvfp4_quantize, and restore the
previous value at the end of the test to avoid leaking global state.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@tests/moe/test_trtllm_gen_fused_moe.py`:
- Around line 2680-2687: run_moe_test currently only uses use_4over6 for
skipping but never actually sets the process env, so 4over6 paths may not be
exercised; wrap the quantize/reference/production section inside the
set_nvfp4_4over6_env context by calling set_nvfp4_4over6_env(use_4over6) (and
ensure the helper is imported) before entering the FP4
quantize/reference/production logic in run_moe_test and restore/unset it after
that block so the FLASHINFER_NVFP4_4OVER6 env state is consistently applied only
for those test cases.

In `@tests/moe/test_trtllm_gen_moe_autotune_tactics.py`:
- Around line 160-169: The test never actually enables the 4over6 NVFP4 runtime
flag because set_nvfp4_4over6_env is never applied; update the test harness so
that when _quant_mode_config is called with use_4over6=True the runtime
environment is toggled for the kernel run: call set_nvfp4_4over6_env(True)
before invoking _run_kernel_with_tactic (and set_nvfp4_4over6_env(False) or
restore the previous state after) so the launched kernel uses the 4over6 path;
adjust every place that constructs the use_4over6=True matrix (including the
other occurrences you noted) to wrap the kernel invocation with the env setter
rather than only changing scales.

In `@tests/moe/test_trtllm_gen_routed_fused_moe.py`:
- Around line 82-83: The test toggles use_4over6 but never actually flips the
NVFP4 4over6 environment, so fp4_quantize() and the routed/non-routed MoE kernel
calls still use the global env; fix by wrapping the sections that perform FP4
quantization and invoke the MoE kernels (references: fp4_quantize, the routed
MoE kernel call(s) and the non-routed MoE kernel call(s)) in the
set_nvfp4_4over6_env context when use_4over6 is True (e.g., with
set_nvfp4_4over6_env(): ...) so the env is applied for those operations and is
restored afterward; apply this same wrapping to the other similar test blocks
currently duplicated later in the file.

In `@tests/moe/utils.py`:
- Around line 40-65: The fixture set_nvfp4_4over6_env currently force-sets
TRTLLM_DISABLE_FP4_QUANT_FAST_MATH and
FLASHINFER_NVFP4_4OVER6_DISABLE_MSE_FAST_MATH unconditionally; change it so
those two env vars are only set when request.getfixturevalue("use_4over6") is
truthy (i.e., set them inside the branch where use_4over6 is True and leave them
untouched when False), while still recording original_values and restoring them
after yield; keep FLASHINFER_NVFP4_4OVER6 set to "1"/"0" based on use_4over6 as
before.

---

Outside diff comments:
In `@tests/moe/test_trtllm_gen_per_token_moe.py`:
- Around line 114-134: The test only updates Python-side scales via
nvfp4_global_decode_scale_te but never flips the backend mode, so wrap the
quantize+kernel calls with the shared helper set_nvfp4_4over6_env(use_4over6) so
the backend is actually in 4over6 mode when calling nvfp4_quantize and
trtllm_fp4_block_scale_routed_moe; specifically, call
set_nvfp4_4over6_env(use_4over6) around the block that computes
hidden_states/hidden_states_scale/per_token_scale_inv with nvfp4_quantize and
the subsequent trtllm_fp4_block_scale_routed_moe invocation so both scale
computation and kernel execution use the same mode (references:
nvfp4_global_decode_scale_te, nvfp4_quantize, set_nvfp4_4over6_env,
trtllm_fp4_block_scale_routed_moe).

---

Duplicate comments:
In `@csrc/nv_internal/cpp/kernels/quantization.cu`:
- Around line 338-362: The code reads the process-global use4Over6
unconditionally which causes FP32 instantiations to abort; fix by making the
env-read T-aware: move or re-evaluate
tensorrt_llm::common::getEnvNVFP4Use4Over6() into the template/lambda scope
where T is visible (the launchKernel capture/instantiation) and force it false
for T=float (e.g. compute auto const use4Over6 =
tensorrt_llm::common::getEnvNVFP4Use4Over6() && !std::is_same_v<T,float> and
pass that as the use4Over6Tag/std::bool_constant), then remove or leave the
now-unreachable TLLM_CHECK_WITH_INFO(!USE_4OVER6, ...) in the float branch.

---

Nitpick comments:
In `@tests/test_helpers/utils_fp4.py`:
- Around line 295-302: The loop computes per-block MSE by accumulating squared
differences across the last dim; replace the explicit for-loop with a vectorized
reduction: compute diff4 = dq4 - x_blocks and diff6 = dq6 - x_blocks, square
them and sum over the last axis to produce err4 and err6, then set pick_four =
err4 < err6 (preserving the strict < tiebreak). Update variables err4, err6,
diff4, diff6 and use the existing dq4, dq6, x_blocks, and pick_four names so the
change is localized to that block.