Let $L=(\mathbf Z^n,b)$ be a bilinear module such that $L\otimes \mathbf Q$ is non-degenerate.
For a set of ultrametric places $S$, let us write $\mathbf Z[S^{-1}]$ for the set of rationals that are integral at each place out of $S$, and $\mathbf Z_{(S)}$ for the set of rationals that are integral at each place in $S$, so that $\mathbf Z[S^{-1}]$ is the same as $\mathbf Z_{(T)}$ when $T$ is the complement of $S$.
Let $\mathrm{Gen}(L)$ be the genus of $L$. One may ask two distinct questions :
(Q1) Does there exist a finite set of ultrametric places $T$ such that for all $M$ in $\mathrm{Gen}(L)$, there exists an isomorphism from $L\otimes \mathbf Z[T^{-1}]$ to $M\otimes \mathbf Z[T^{-1}]$ ?
(Q2) What is the biggest subset $S$ of the set of ultrametric places such that for all $M$ in $\mathrm{Gen}(L)$, one has an equivalence between the two affirmations $L\otimes \mathbf Z[S^{-1}]\simeq M\otimes \mathbf Z[S^{-1}]$ and $L\simeq M$.
In my first answer, I assumed you were asking for (Q1). I am not sure of this anymore. So here are answers to both questions.
(Q1) : This is an application of the strong approximation theorem.
Recall that it says that if $T$ is a finite set of places containing the archimedean ones, and if $S$ is its complement in the set of all places, and if $L$ is isotropic at at least one place in $T$, then the spinor genus of $L\otimes \mathbf Z_{(S)}$ contains a single class. Now at least when the dimension exceeds $3$, the set of ultrametric places $p$ where the equivalence relation defining spinor genera differs from the equivalence relation defining genera is finite (the two equivalence relations at the place $p$ agree when the determinant is a unit in $\mathbf Z_p$). Thus it suffices to choose $T$ big enough to contain all of these "bad" classes, and also a place where the form is isotropic.
(Q2) : This biggest set can be empty. Here is an example : take for $(L,q)$ the sum $I_9$ of nine copies of $I_1:=(\mathbf Z,(x,y)\mapsto xy)$. There are two classes in the genus of $L$; the one is represented by $L$, the other is represented by $M:=\mathrm{E}_8\oplus I_1$. It happens that $M\otimes\mathbf Z[\frac{1}{p}]$ is isomorphic to $L\otimes\mathbf Z[\frac{1}{p}]$ for all prime $p$.
It can also be non-empty : For $I_n$ with $n\leq 8$, the set $S$ is the set of all ultrametric places, since then the genus of $I_n$ contains a single class.
These two cases are representative of the general situation : if the genus of $L$ contains a unique class, then $S$ consists of all places. If the spinor genus contains more than one class, then $S$ cannot contain any place where $L$ is isotropic, and thus is empty when the dimension exceeds 5.
The canonical reference for this kind of questions is O'Meara.
Addendum (4/12/13): In dimensions $2$, funny things happen : by a theorem of Gauss, the theory (at least on the level of proper genera) is the same as the theory of narrow class groups $C(O)$ of orders in quadratic extensions of $\mathbf Q$. That question $(Q1)$ admits a positive answer then follows from the fact that these class groups are finitely generated (and in fact even finite). And $(Q_2)$ asks for an order $O$ what is the set $S$ of primes $p$ such that the ideals dividing $pO$ are trivial in $C(O)$. I guess that the proportion of these primes in $S$ in $\frac{1}{\vert C(O) \vert}$ but might be wrong on this point.
